How is pricing calculated?

Pricing is based on the hourly rate of operators in your selected region. You only pay for the time operators spend controlling your robots.

What robots are supported?

We support 9+ robot arms including SO-100, Koch, Franka FR3, Franka Panda, WidowX 250, ViperX 300, ViperX 300s, and ALOHA.

What is the latency for teleoperation?

Our platform provides ultra-low latency of 30-70ms worldwide, enabling real-time responsive robot control.

What is a vision language action model?

A vision language action model, usually shortened to VLA, takes camera images and a task written in plain language and outputs robot actions, normally joint positions or end effector deltas. The vision and language part is typically a pretrained vision language model, and an action head converts its representation into continuous commands. The point of the design is that a single policy can follow instructions it was never explicitly trained on, instead of one policy per task.

Why does the arena not rank models with one overall score?

Because the numbers come from different benchmark suites, and a LIBERO success rate and a real robot success rate do not measure the same thing. Averaging them would produce a number that looks authoritative and means nothing. The arena therefore sorts inside a column, always shows which suite a value came from, and links every value to the source it was taken from.

How much GPU memory does a VLA need?

Most papers do not publish a memory figure, so the arena computes the weights only footprint from the parameter count at bf16 precision and marks it as computed. Weights are a lower bound: activations, the image encoder intermediates, the KV cache and the CUDA context come on top, which in practice is roughly 30 to 50 percent more for inference. Fine tuning is a different order of magnitude again, because optimiser state and gradients usually cost several times the weights unless you use LoRA or a similar adapter.

Which of these models can I actually train on AY-Robots?

Five of them: GR00T N1.7, GR00T N1.5, Pi0.5, SmolVLA and ACT. Those are the models the training pool can fine tune on a LeRobot dataset that you record with your own arm, and the platform pages for them carry the real defaults, GPU tier and measured inference latency from that pool. Every other model in the arena is listed for comparison and links to its own repository.

What does the community rating mean?

It is a one to five rating from people with an AY-Robots account, one vote per account and per model, changeable at any time. It is a subjective signal about how a model behaved in practice, not a benchmark. Read it next to the measured numbers, not instead of them.

How current is this list?

Every entry carries its own editorial date and the sources it was built from. The field moves fast, so a model can gain a new checkpoint or a new licence between updates. Where a source contradicts what you read here, trust the source and tell us, the link is on every model page.

VLA Arena

Every vision language action model, side by side

Parameters, GPU memory, inference latency, licence and what each model actually scored in simulation, on real hardware and after fine tuning. Every value links to the paper, model card or repository it came from. Nothing here is estimated unless it says so.

85 models
57 with open weights
32 developers
332 sourced results

Open the table Models you can train here

The table

Sort by any column, filter by category, licence or whether this platform can train the model for you. GPU memory is the weights only footprint at bf16 computed from the parameter count, unless a published figure exists, in which case that one is shown and marked as reported.

Showing 85 of 85 models. Every number links to the source on the model page. Values from different benchmark suites are not comparable, so sort inside one column rather than across columns.

ACE
ACE-Brain-0.5
ACE-Brain, July 2026
Parameters
8.8 B
GPU memory
16 GB weights, 24 GB card
Latency
not published
Weights
see model card
Simulation
98.2% LIBERO
Real world
86.3% MindCube
Gemini Robotics 2
Google DeepMind, July 2026
Parameters
not published
GPU memory
not published
Latency
not published
Weights
closed
Real world
92% Multi-finger dexterity (Google DeepMind internal)
Gemini Robotics ER 2
Google DeepMind, July 2026
Parameters
not published
GPU memory
not published
Latency
not published
Weights
closed
Gemini Robotics On-Device 2
Google DeepMind, July 2026
Parameters
not published
GPU memory
not published
Latency
not published
Weights
closed
RBY
LingBot-VLA 2.0
Robbyant, July 2026
Parameters
6 B
GPU memory
11 GB weights, 16 GB card
Latency
130 ms
Weights
Apache-2.0
Simulation
93.52% RoboTwin 2.0
Real world
66.2% GM-100
AI2
MolmoAct2
Allen Institute for AI, May 2026
Parameters
5.5 B
GPU memory
10 GB weights, 16 GB card
Latency
180 ms
Weights
Apache-2.0
Simulation
100% LIBERO
Real world
44.3% RoboEval
AI2
MolmoAct2-Think
Allen Institute for AI, May 2026
Parameters
5.5 B
GPU memory
10 GB weights, 16 GB card
Latency
790 ms
Weights
Apache-2.0
Simulation
98.8% LIBERO
XSR
Wall-OSS-0.5
X Square Robot, May 2026
Parameters
4 B
GPU memory
7.5 GB weights, 12 GB card
Latency
not published
Weights
Apache-2.0
Simulation
96.5% LIBERO
Real world
61.1% Real robot, 15-task suite (Wall-OSS-0.5 report)
Gemini Robotics-ER 1.6
Google DeepMind, April 2026
Parameters
not published
GPU memory
not published
Latency
not published
Weights
closed
Real world
93% Instrument reading (Google DeepMind internal)
GEN
GEN-1
Generalist AI, April 2026
Parameters
not published
GPU memory
not published
Latency
not published
Weights
closed
Real world
99% Generalist AI internal real-robot evaluation
AGI
GO-2 (Genie Operator-2)
AgiBot, April 2026
Parameters
not published
GPU memory
not published
Latency
not published
Weights
closed
Simulation
98.5% LIBERO
NVIDIA Isaac GR00T N1.7
NVIDIA, April 2026
Parameters
3.1 B
GPU memory
16 GB reported
Latency
27.9 ms
Weights
NVIDIA Open Model
Simulation
98.45% LIBERO
Trainable here
π
π0.7
Physical Intelligence, April 2026
Parameters
5 B
GPU memory
9.3 GB weights, 16 GB card
Latency
not published
Weights
closed
Cosmos Policy
NVIDIA, January 2026
Parameters
2 B
GPU memory
6.8 GB reported
Latency
not published
Weights
Code Apache-2.0
Simulation
98.5% LIBERO
FIG
Helix 02
Figure AI, January 2026
Parameters
not published
GPU memory
not published
Latency
not published
Weights
closed
RBY
LingBot-VLA
Robbyant, January 2026
Parameters
4 B
GPU memory
7.5 GB weights, 12 GB card
Latency
not published
Weights
Apache-2.0
Simulation
88.56% RoboTwin 2.0
Real world
35.41% GM-100
UNI
UnifoLM-VLA-0
Unitree, January 2026
Parameters
not published
GPU memory
not published
Latency
not published
Weights
CC BY-NC-SA
Simulation
100% LIBERO
NVIDIA Isaac GR00T N1.6
NVIDIA, December 2025
Parameters
3.3 B
GPU memory
6.1 GB weights, 12 GB card
Latency
44 ms
Weights
NVIDIA OneWay NC
Simulation
98.45% LIBERO
SUT
NORA-1.5
SUTD, November 2025
Parameters
4 B
GPU memory
7.5 GB weights, 12 GB card
Latency
not published
Weights
MIT
Simulation
95% LIBERO
RynnVLA-002
Alibaba, November 2025
Parameters
not published
GPU memory
not published
Latency
not published
Weights
Apache-2.0
Simulation
97.4% LIBERO
Real world
50% Real robot (LeRobot SO100)
π
π*0.6
Physical Intelligence, November 2025
Parameters
not published
GPU memory
not published
Latency
not published
Weights
closed
Real world
90% Physical Intelligence real-robot RECAP evaluation
SAI
InternVLA-M1
Shanghai AI Laboratory, October 2025
Parameters
4.1 B
GPU memory
12 GB reported
Latency
not published
Weights
Code MIT
Simulation
95.9% LIBERO
THU
X-VLA-0.9B
Tsinghua University, October 2025
Parameters
880 M
GPU memory
1.6 GB weights, 8 GB card
Latency
not published
Weights
Apache-2.0
Simulation
98.1% LIBERO
Gemini Robotics 1.5
Google DeepMind, September 2025
Parameters
not published
GPU memory
not published
Latency
not published
Weights
closed
Gemini Robotics-ER 1.5
Google DeepMind, September 2025
Parameters
not published
GPU memory
not published
Latency
not published
Weights
closed
THU
RDT2-FM
Tsinghua University, September 2025
Parameters
not published
GPU memory
16 GB reported
Latency
not published
Weights
Apache-2.0
THU
RDT2-VQ
Tsinghua University, September 2025
Parameters
8.3 B
GPU memory
16 GB reported
Latency
not published
Weights
Apache-2.0
XSR
WALL-OSS (flow)
X Square Robot, September 2025
Parameters
4 B
GPU memory
7.5 GB weights, 12 GB card
Latency
not published
Weights
Apache-2.0
XSR
WALL-OSS-FAST
X Square Robot, September 2025
Parameters
not published
GPU memory
not published
Latency
not published
Weights
Apache-2.0
AI2
MolmoAct-7B-D
Allen Institute for AI, August 2025
Parameters
8.1 B
GPU memory
15 GB weights, 24 GB card
Latency
not published
Weights
Apache-2.0
Simulation
95.4% LIBERO
AI2
MolmoAct-7B-O
Allen Institute for AI, August 2025
Parameters
7.7 B
GPU memory
14 GB weights, 24 GB card
Latency
not published
Weights
Apache-2.0
RynnVLA-001
Alibaba, August 2025
Parameters
7 B
GPU memory
13 GB weights, 24 GB card
Latency
not published
Weights
Apache-2.0
Real world
91.7% Real robot (LeRobot SO-100)
GR-3
ByteDance, July 2025
Parameters
4 B
GPU memory
7.5 GB weights, 12 GB card
Latency
not published
Weights
closed
Real world
97.5% Real robot (ByteMini), table bussing (long-horizon)
villa-X
Microsoft, July 2025
Parameters
not published
GPU memory
not published
Latency
not published
Weights
closed
Simulation
90.1% LIBERO
CAS
BitVLA
Chinese Academy of Sciences, June 2025
Parameters
3 B
GPU memory
1.4 GB reported
Latency
73 ms
Weights
MIT
Simulation
99% LIBERO
Gemini Robotics On-Device
Google DeepMind, June 2025
Parameters
not published
GPU memory
not published
Latency
not published
Weights
closed
SmolVLA
Hugging Face, June 2025
Parameters
450 M
GPU memory
0.8 GB weights, 8 GB card
Latency
not published
Weights
Apache-2.0
Simulation
87.3% LIBERO
Real world
78.3% SO-100 real robot
Trainable here
BAI
UniVLA (Unified Vision-Language-Action Model, BAAI)
BAAI, June 2025
Parameters
8.5 B
GPU memory
16 GB weights, 24 GB card
Latency
not published
Weights
Apache-2.0
Simulation
95.5% LIBERO
MID
ChatVLA-2
Midea Group, May 2025
Parameters
not published
GPU memory
not published
Latency
not published
Weights
closed
NVIDIA Isaac GR00T N1.5
NVIDIA, May 2025
Parameters
2.7 B
GPU memory
5.1 GB weights, 8 GB card
Latency
not published
Weights
NVIDIA OneWay NC
Simulation
47.5% RoboCasa
Real world
93.3% Real GR-1 humanoid manipulation, language following
Trainable here
ODL
UniVLA (task-centric latent actions)
OpenDriveLab, May 2025
Parameters
7.5 B
GPU memory
14 GB weights, 24 GB card
Latency
not published
Weights
Apache-2.0
Simulation
95.2% LIBERO
Real world
75% Real robot (UniVLA paper setup)
SUT
NORA
SUTD, April 2025
Parameters
3 B
GPU memory
8.3 GB reported
Latency
not published
Weights
MIT
Simulation
87.9% LIBERO
Real world
56.7% Real-world WidowX (authors' setup)
π
π0.5
Physical Intelligence, April 2025
Parameters
3.6 B
GPU memory
8 GB reported
Latency
not published
Weights
Apache-2.0
Simulation
98.8% LIBERO
Trainable here
Gemini Robotics
Google DeepMind, March 2025
Parameters
not published
GPU memory
not published
Latency
250 ms
Weights
closed
Real world
100% Gemini Robotics dexterous long-horizon specialist tasks
Gemini Robotics-ER
Google DeepMind, March 2025
Parameters
not published
GPU memory
not published
Latency
not published
Weights
closed
Simulation
65% ALOHA 2 sim task suite
Real world
65% ALOHA 2 real robot
AGI
GO-1 (Genie Operator-1)
AgiBot, March 2025
Parameters
3 B
GPU memory
5.6 GB weights, 8 GB card
Latency
not published
Weights
CC BY-NC-SA
Real world
60% Real robot (complex long-horizon and dexterous tasks)
NVIDIA Isaac GR00T N1
NVIDIA, March 2025
Parameters
2.2 B
GPU memory
4.1 GB weights, 8 GB card
Latency
63.9 ms
Weights
NVIDIA OneWay NC
Simulation
66.5% DexMimicGen (DexMG)
Real world
76.8% Real-world GR-1 humanoid evaluation
MID
ChatVLA
Midea Group, February 2025
Parameters
3.4 B
GPU memory
6.3 GB weights, 12 GB card
Latency
not published
Weights
MIT
Real world
51.4% Real robot (authors' setup)
MID
DexVLA
Midea Group, February 2025
Parameters
not published
GPU memory
not published
Latency
not published
Weights
closed
Simulation
99.1% LIBERO
FIG
Helix
Figure AI, February 2025
Parameters
7.1 B
GPU memory
13 GB weights, 24 GB card
Latency
not published
Weights
closed
Real world
95% Figure internal logistics evaluation (parcel induction)
Magma-8B
Microsoft, February 2025
Parameters
8.9 B
GPU memory
17 GB reported
Latency
1100 ms
Weights
MIT
Simulation
52.3% SimplerEnv
Real world
67.3% AITW
SU
OpenVLA-OFT
Stanford University, February 2025
Parameters
7.5 B
GPU memory
14 GB weights, 24 GB card
Latency
72.9 ms
Weights
MIT
Simulation
97.1% LIBERO
SAI
SpatialVLA
Shanghai AI Laboratory, January 2025
Parameters
3.5 B
GPU memory
8.5 GB reported
Latency
not published
Weights
MIT
Simulation
78.1% LIBERO
π
π0-FAST
Physical Intelligence, January 2025
Parameters
3 B
GPU memory
8 GB reported
Latency
750 ms
Weights
Apache-2.0
HKU
Moto (Moto-GPT)
HKU, December 2024
Parameters
98 M
GPU memory
0.2 GB weights, 8 GB card
Latency
not published
Weights
Apache-2.0
Simulation
74% SimplerEnv
Real world
60% Real-world (authors' own three-task suite)
SAI
Seer
Shanghai AI Laboratory, December 2024
Parameters
316 M
GPU memory
0.6 GB weights, 8 GB card
Latency
not published
Weights
closed
Simulation
87.7% LIBERO
Real world
78.4% Real-world (authors' own four-task suite)
SAI
Seer-Large
Shanghai AI Laboratory, December 2024
Parameters
566 M
GPU memory
1.1 GB weights, 8 GB card
Latency
not published
Weights
closed
CogACT-Base
Microsoft, November 2024
Parameters
7.6 B
GPU memory
30 GB reported
Latency
181 ms
Weights
MIT
Simulation
74.8% SIMPLER
Real world
71.2% Real robot (Realman arm)
CogACT-Large
Microsoft, November 2024
Parameters
not published
GPU memory
30 GB reported
Latency
not published
Weights
MIT
Simulation
76.7% SIMPLER
CogACT-Small
Microsoft, November 2024
Parameters
not published
GPU memory
30 GB reported
Latency
not published
Weights
MIT
Simulation
73.3% SIMPLER
GR-2
ByteDance, October 2024
Parameters
230 M
GPU memory
0.4 GB weights, 8 GB card
Latency
not published
Weights
closed
Real world
97.7% Real robot, multi-task (ByteDance platform)
KST
LAPA (Latent Action Pretraining)
KAIST, October 2024
Parameters
7 B
GPU memory
13 GB weights, 24 GB card
Latency
not published
Weights
MIT
Simulation
62% Language Table (simulation)
Real world
50.1% Real-world tabletop manipulation (Franka Emika Panda)
THU
RDT-170M
Tsinghua University, October 2024
Parameters
170 M
GPU memory
0.3 GB weights, 8 GB card
Latency
not published
Weights
MIT
THU
RDT-1B
Tsinghua University, October 2024
Parameters
1.2 B
GPU memory
2.2 GB weights, 8 GB card
Latency
not published
Weights
MIT
Real world
100% Real robot (Mobile ALOHA, RDT paper protocol)
π
π0 (pi-zero)
Physical Intelligence, October 2024
Parameters
3.3 B
GPU memory
8 GB reported
Latency
73 ms
Weights
Apache-2.0
MIT
HPT-Base
MIT, September 2024
Parameters
13 M
GPU memory
0 GB weights, 8 GB card
Latency
not published
Weights
MIT
Real world
70% Real-world (authors' own setup)
MIT
HPT-Large
MIT, September 2024
Parameters
51 M
GPU memory
0.1 GB weights, 8 GB card
Latency
not published
Weights
MIT
MIT
HPT-Small
MIT, September 2024
Parameters
3 M
GPU memory
0 GB weights, 8 GB card
Latency
not published
Weights
MIT
MIT
HPT-XLarge
MIT, September 2024
Parameters
227 M
GPU memory
0.4 GB weights, 8 GB card
Latency
not published
Weights
MIT
Real world
76.7% Real-world (authors' own setup)
MID
TinyVLA
Midea Group, September 2024
Parameters
1.3 B
GPU memory
2.4 GB weights, 8 GB card
Latency
14 ms
Weights
MIT
Simulation
77.6% Meta-World
Real world
94% Real-world Franka (authors' setup)
UCB
CrossFormer
UC Berkeley, August 2024
Parameters
130 M
GPU memory
0.2 GB weights, 8 GB card
Latency
not published
Weights
MIT
Real world
95% CrossFormer real-robot evaluation (paper Table 3)
UCB
Embodied Chain-of-Thought (ECoT)
UC Berkeley, July 2024
Parameters
7.2 B
GPU memory
13 GB weights, 24 GB card
Latency
not published
Weights
MIT
Real world
72% ECoT accelerated inference study
SU
OpenVLA
Stanford University, June 2024
Parameters
7.5 B
GPU memory
16.8 GB reported
Latency
239.6 ms
Weights
MIT
Simulation
76.5% LIBERO
Real world
85% Google robot (real RT-1 mobile manipulator)
PKU
RoboMamba
Peking University, June 2024
Parameters
3.2 B
GPU memory
6 GB weights, 12 GB card
Latency
not published
Weights
closed
Simulation
63% SAPIEN manipulation
UCB
Octo-Base
UC Berkeley, May 2024
Parameters
93 M
GPU memory
0.2 GB weights, 8 GB card
Latency
not published
Weights
MIT
Simulation
75.1% LIBERO
Real world
72% Octo real-robot finetuning suite (paper Table I)
UCB
Octo-Small
UC Berkeley, May 2024
Parameters
27 M
GPU memory
0.1 GB weights, 8 GB card
Latency
not published
Weights
MIT
UMA
3D-VLA
UMass Amherst, March 2024
Parameters
not published
GPU memory
not published
Latency
not published
Weights
closed
Simulation
68% RLBench
OXE
RT-1-X
Open X-Embodiment, October 2023
Parameters
not published
GPU memory
not published
Latency
not published
Weights
Apache-2.0
Real world
50% Open X-Embodiment cross-lab evaluation
OXE
RT-2-X
Open X-Embodiment, October 2023
Parameters
55 B
GPU memory
102 GB weights
Latency
not published
Weights
closed
RT-2
Google DeepMind, July 2023
Parameters
55 B
GPU memory
102 GB weights
Latency
not published
Weights
closed
Simulation
90% Language-Table (simulation)
RoboCat
Google DeepMind, June 2023
Parameters
1.2 B
GPU memory
2.2 GB weights, 8 GB card
Latency
not published
Weights
closed
SU
ACT (Action Chunking with Transformers)
Stanford University, April 2023
Parameters
80 M
GPU memory
0.1 GB weights, 8 GB card
Latency
10 ms
Weights
MIT
Simulation
86% ALOHA simulation (MuJoCo)
Real world
96% Real ALOHA hardware
Trainable here
CU
Diffusion Policy (CNN-based, DiffusionPolicy-C)
Columbia University, March 2023
Parameters
278 M
GPU memory
0.5 GB weights, 8 GB card
Latency
100 ms
Weights
MIT
Simulation
93% Robomimic
CU
Diffusion Policy (Transformer-based, DiffusionPolicy-T)
Columbia University, March 2023
Parameters
31 M
GPU memory
0.1 GB weights, 8 GB card
Latency
not published
Weights
closed
Simulation
90% Robomimic
RT-1
Google, December 2022
Parameters
35 M
GPU memory
0.1 GB weights, 8 GB card
Latency
not published
Weights
Apache-2.0
Real world
97% RT-1 real-robot evaluation


ACE ACE-Brain-0.5 Reasoning VLA	ACE-Brain	July 2026	8.8 B see model card	16 GB weights, 24 GB card	not published	98.2%LIBERO	86.3%MindCube	not reported	Be the first
Gemini Robotics 2 Foundation VLA	Google DeepMind	July 2026	not published closed weights	not published	not published	not reported	92%Multi-finger dexterity (Google DeepMind internal)	not reported	Be the first
Gemini Robotics ER 2 Reasoning VLA	Google DeepMind	July 2026	not published closed weights	not published	not published	not reported	not reported	not reported	Be the first
Gemini Robotics On-Device 2 Compact VLA	Google DeepMind	July 2026	not published closed weights	not published	not published	not reported	not reported	not reported	Be the first
RBY LingBot-VLA 2.0 Foundation VLA	Robbyant	July 2026	6 B Apache-2.0	11 GB weights, 16 GB card	130 ms	93.52%RoboTwin 2.0	66.2%GM-100	not reported	Be the first
AI2 MolmoAct2 Reasoning VLA	Allen Institute for AI	May 2026	5.5 B Apache-2.0	10 GB weights, 16 GB card	180 ms	100%LIBERO	44.3%RoboEval	87.1%Real-world zero-shot (Franka DROID setup)	Be the first
AI2 MolmoAct2-Think Reasoning VLA	Allen Institute for AI	May 2026	5.5 B Apache-2.0	10 GB weights, 16 GB card	790 ms	98.8%LIBERO	not reported	not reported	Be the first
XSR Wall-OSS-0.5 Foundation VLA	X Square Robot	May 2026	4 B Apache-2.0	7.5 GB weights, 12 GB card	not published	96.5%LIBERO	61.1%Real robot, 15-task suite (Wall-OSS-0.5 report)	51.1%Real robot, 17-task zero-shot suite (Wall-OSS-0.5 report)	Be the first
Gemini Robotics-ER 1.6 Reasoning VLA	Google DeepMind	April 2026	not published closed weights	not published	not published	not reported	93%Instrument reading (Google DeepMind internal)	not reported	Be the first
GEN GEN-1 Foundation VLA	Generalist AI	April 2026	not published closed weights	not published	not published	not reported	99%Generalist AI internal real-robot evaluation	not reported	Be the first
AGI GO-2 (Genie Operator-2) Reasoning VLA	AgiBot	April 2026	not published closed weights	not published	not published	98.5%LIBERO	not reported	not reported	Be the first
NVIDIA Isaac GR00T N1.7 Reasoning VLA Trainable here	NVIDIA	April 2026	3.1 B NVIDIA Open Model	16 GB reported	27.9 ms	98.45%LIBERO	not reported	not reported	Be the first
π π0.7 Foundation VLA	Physical Intelligence	April 2026	5 B closed weights	9.3 GB weights, 16 GB card	not published	not reported	not reported	85.6%Physical Intelligence real-robot evaluation	Be the first
Cosmos Policy World model	NVIDIA	January 2026	2 B Code Apache-2.0	6.8 GB reported	not published	98.5%LIBERO	not reported	not reported	Be the first
FIG Helix 02 Foundation VLA	Figure AI	January 2026	not published closed weights	not published	not published	not reported	not reported	not reported	Be the first
RBY LingBot-VLA Foundation VLA	Robbyant	January 2026	4 B Apache-2.0	7.5 GB weights, 12 GB card	not published	88.56%RoboTwin 2.0	35.41%GM-100	not reported	Be the first
UNI UnifoLM-VLA-0 Foundation VLA	Unitree	January 2026	not published CC BY-NC-SA	not published	not published	100%LIBERO	not reported	not reported	Be the first
NVIDIA Isaac GR00T N1.6 Reasoning VLA	NVIDIA	December 2025	3.3 B NVIDIA OneWay NC	6.1 GB weights, 12 GB card	44 ms	98.45%LIBERO	not reported	not reported	Be the first
SUT NORA-1.5 Compact VLA	SUTD	November 2025	4 B MIT	7.5 GB weights, 12 GB card	not published	95%LIBERO	not reported	not reported	Be the first
RynnVLA-002 World model	Alibaba	November 2025	not published Apache-2.0	not published	not published	97.4%LIBERO	50%Real robot (LeRobot SO100)	not reported	Be the first
π π*0.6 Reasoning VLA	Physical Intelligence	November 2025	not published closed weights	not published	not published	not reported	90%Physical Intelligence real-robot RECAP evaluation	not reported	Be the first
SAI InternVLA-M1 Foundation VLA	Shanghai AI Laboratory	October 2025	4.1 B Code MIT	12 GB reported	not published	95.9%LIBERO	not reported	not reported	Be the first
THU X-VLA-0.9B Foundation VLA	Tsinghua University	October 2025	880 M Apache-2.0	1.6 GB weights, 8 GB card	not published	98.1%LIBERO	not reported	not reported	Be the first
Gemini Robotics 1.5 Foundation VLA	Google DeepMind	September 2025	not published closed weights	not published	not published	not reported	not reported	not reported	Be the first
Gemini Robotics-ER 1.5 Reasoning VLA	Google DeepMind	September 2025	not published closed weights	not published	not published	not reported	not reported	not reported	Be the first
THU RDT2-FM Foundation VLA	Tsinghua University	September 2025	not published Apache-2.0	16 GB reported	not published	not reported	not reported	not reported	Be the first
THU RDT2-VQ Foundation VLA	Tsinghua University	September 2025	8.3 B Apache-2.0	16 GB reported	not published	not reported	not reported	not reported	Be the first
XSR WALL-OSS (flow) Foundation VLA	X Square Robot	September 2025	4 B Apache-2.0	7.5 GB weights, 12 GB card	not published	not reported	not reported	not reported	Be the first
XSR WALL-OSS-FAST Foundation VLA	X Square Robot	September 2025	not published Apache-2.0	not published	not published	not reported	not reported	not reported	Be the first
AI2 MolmoAct-7B-D Reasoning VLA	Allen Institute for AI	August 2025	8.1 B Apache-2.0	15 GB weights, 24 GB card	not published	95.4%LIBERO	not reported	not reported	Be the first
AI2 MolmoAct-7B-O Reasoning VLA	Allen Institute for AI	August 2025	7.7 B Apache-2.0	14 GB weights, 24 GB card	not published	not reported	not reported	not reported	Be the first
RynnVLA-001 Foundation VLA	Alibaba	August 2025	7 B Apache-2.0	13 GB weights, 24 GB card	not published	not reported	91.7%Real robot (LeRobot SO-100)	not reported	Be the first
GR-3 Foundation VLA	ByteDance	July 2025	4 B closed weights	7.5 GB weights, 12 GB card	not published	not reported	97.5%Real robot (ByteMini), table bussing (long-horizon)	not reported	Be the first
villa-X Foundation VLA	Microsoft	July 2025	not published closed weights	not published	not published	90.1%LIBERO	not reported	not reported	Be the first
CAS BitVLA Compact VLA	Chinese Academy of Sciences	June 2025	3 B MIT	1.4 GB reported	73 ms	99%LIBERO	not reported	not reported	Be the first
Gemini Robotics On-Device Compact VLA	Google DeepMind	June 2025	not published closed weights	not published	not published	not reported	not reported	not reported	Be the first
SmolVLA Compact VLA Trainable here	Hugging Face	June 2025	450 M Apache-2.0	0.8 GB weights, 8 GB card	not published	87.3%LIBERO	78.3%SO-100 real robot	not reported	Be the first
BAI UniVLA (Unified Vision-Language-Action Model, BAAI) Foundation VLA	BAAI	June 2025	8.5 B Apache-2.0	16 GB weights, 24 GB card	not published	95.5%LIBERO	not reported	not reported	Be the first
MID ChatVLA-2 Reasoning VLA	Midea Group	May 2025	not published closed weights	not published	not published	not reported	not reported	not reported	Be the first
NVIDIA Isaac GR00T N1.5 Foundation VLA Trainable here	NVIDIA	May 2025	2.7 B NVIDIA OneWay NC	5.1 GB weights, 8 GB card	not published	47.5%RoboCasa	93.3%Real GR-1 humanoid manipulation, language following	98.8%Unitree G1 post-training	Be the first
ODL UniVLA (task-centric latent actions) Foundation VLA	OpenDriveLab	May 2025	7.5 B Apache-2.0	14 GB weights, 24 GB card	not published	95.2%LIBERO	75%Real robot (UniVLA paper setup)	not reported	Be the first
SUT NORA Compact VLA	SUTD	April 2025	3 B MIT	8.3 GB reported	not published	87.9%LIBERO	56.7%Real-world WidowX (authors' setup)	not reported	Be the first
π π0.5 Foundation VLA Trainable here	Physical Intelligence	April 2025	3.6 B Apache-2.0	8 GB reported	not published	98.8%LIBERO	not reported	not reported	Be the first
Gemini Robotics Foundation VLA	Google DeepMind	March 2025	not published closed weights	not published	250 ms	not reported	100%Gemini Robotics dexterous long-horizon specialist tasks	63%Gemini Robotics embodiment adaptation	Be the first
Gemini Robotics-ER Reasoning VLA	Google DeepMind	March 2025	not published closed weights	not published	not published	65%ALOHA 2 sim task suite	65%ALOHA 2 real robot	25%ALOHA 2 real robot	Be the first
AGI GO-1 (Genie Operator-1) Foundation VLA	AgiBot	March 2025	3 B CC BY-NC-SA	5.6 GB weights, 8 GB card	not published	not reported	60%Real robot (complex long-horizon and dexterous tasks)	not reported	Be the first
NVIDIA Isaac GR00T N1 Foundation VLA	NVIDIA	March 2025	2.2 B NVIDIA OneWay NC	4.1 GB weights, 8 GB card	63.9 ms	66.5%DexMimicGen (DexMG)	76.8%Real-world GR-1 humanoid evaluation	not reported	Be the first
MID ChatVLA Reasoning VLA	Midea Group	February 2025	3.4 B MIT	6.3 GB weights, 12 GB card	not published	not reported	51.4%Real robot (authors' setup)	not reported	Be the first
MID DexVLA Compact VLA	Midea Group	February 2025	not published closed weights	not published	not published	99.1%LIBERO	not reported	not reported	Be the first
FIG Helix Foundation VLA	Figure AI	February 2025	7.1 B closed weights	13 GB weights, 24 GB card	not published	not reported	95%Figure internal logistics evaluation (parcel induction)	not reported	Be the first
Magma-8B Foundation VLA	Microsoft	February 2025	8.9 B MIT	17 GB reported	1100 ms	52.3%SimplerEnv	67.3%AITW	not reported	Be the first
SU OpenVLA-OFT Foundation VLA	Stanford University	February 2025	7.5 B MIT	14 GB weights, 24 GB card	72.9 ms	97.1%LIBERO	not reported	not reported	Be the first
SAI SpatialVLA Foundation VLA	Shanghai AI Laboratory	January 2025	3.5 B MIT	8.5 GB reported	not published	78.1%LIBERO	not reported	11.1%Real-world WidowX (NORA authors' setup)	Be the first
π π0-FAST Foundation VLA	Physical Intelligence	January 2025	3 B Apache-2.0	8 GB reported	750 ms	not reported	not reported	not reported	Be the first
HKU Moto (Moto-GPT) Action policy	HKU	December 2024	98 M Apache-2.0	0.2 GB weights, 8 GB card	not published	74%SimplerEnv	60%Real-world (authors' own three-task suite)	not reported	Be the first
SAI Seer Action policy	Shanghai AI Laboratory	December 2024	316 M closed weights	0.6 GB weights, 8 GB card	not published	87.7%LIBERO	78.4%Real-world (authors' own four-task suite)	not reported	Be the first
SAI Seer-Large Action policy	Shanghai AI Laboratory	December 2024	566 M closed weights	1.1 GB weights, 8 GB card	not published	not reported	not reported	not reported	Be the first
CogACT-Base Foundation VLA	Microsoft	November 2024	7.6 B MIT	30 GB reported	181 ms	74.8%SIMPLER	71.2%Real robot (Realman arm)	not reported	Be the first
CogACT-Large Foundation VLA	Microsoft	November 2024	not published MIT	30 GB reported	not published	76.7%SIMPLER	not reported	not reported	Be the first
CogACT-Small Foundation VLA	Microsoft	November 2024	not published MIT	30 GB reported	not published	73.3%SIMPLER	not reported	not reported	Be the first
GR-2 World model	ByteDance	October 2024	230 M closed weights	0.4 GB weights, 8 GB card	not published	not reported	97.7%Real robot, multi-task (ByteDance platform)	not reported	Be the first
KST LAPA (Latent Action Pretraining) Foundation VLA	KAIST	October 2024	7 B MIT	13 GB weights, 24 GB card	not published	62%Language Table (simulation)	50.1%Real-world tabletop manipulation (Franka Emika Panda)	not reported	Be the first
THU RDT-170M Compact VLA	Tsinghua University	October 2024	170 M MIT	0.3 GB weights, 8 GB card	not published	not reported	not reported	not reported	Be the first
THU RDT-1B Foundation VLA	Tsinghua University	October 2024	1.2 B MIT	2.2 GB weights, 8 GB card	not published	not reported	100%Real robot (Mobile ALOHA, RDT paper protocol)	100%Real robot (Mobile ALOHA, RDT paper protocol)	Be the first
π π0 (pi-zero) Foundation VLA	Physical Intelligence	October 2024	3.3 B Apache-2.0	8 GB reported	73 ms	not reported	not reported	not reported	Be the first
MIT HPT-Base Action policy	MIT	September 2024	13 M MIT	0 GB weights, 8 GB card	not published	not reported	70%Real-world (authors' own setup)	not reported	Be the first
MIT HPT-Large Action policy	MIT	September 2024	51 M MIT	0.1 GB weights, 8 GB card	not published	not reported	not reported	not reported	Be the first
MIT HPT-Small Action policy	MIT	September 2024	3 M MIT	0 GB weights, 8 GB card	not published	not reported	not reported	not reported	Be the first
MIT HPT-XLarge Action policy	MIT	September 2024	227 M MIT	0.4 GB weights, 8 GB card	not published	not reported	76.7%Real-world (authors' own setup)	not reported	Be the first
MID TinyVLA Compact VLA	Midea Group	September 2024	1.3 B MIT	2.4 GB weights, 8 GB card	14 ms	77.6%Meta-World	94%Real-world Franka (authors' setup)	not reported	Be the first
UCB CrossFormer Action policy	UC Berkeley	August 2024	130 M MIT	0.2 GB weights, 8 GB card	not published	not reported	95%CrossFormer real-robot evaluation (paper Table 3)	82%CrossFormer real-robot evaluation (paper Table 3)	Be the first
UCB Embodied Chain-of-Thought (ECoT) Reasoning VLA	UC Berkeley	July 2024	7.2 B MIT	13 GB weights, 24 GB card	not published	not reported	72%ECoT accelerated inference study	not reported	Be the first
SU OpenVLA Foundation VLA	Stanford University	June 2024	7.5 B MIT	16.8 GB reported	239.6 ms	76.5%LIBERO	85%Google robot (real RT-1 mobile manipulator)	not reported	Be the first
PKU RoboMamba Reasoning VLA	Peking University	June 2024	3.2 B closed weights	6 GB weights, 12 GB card	not published	63%SAPIEN manipulation	not reported	not reported	Be the first
UCB Octo-Base Action policy	UC Berkeley	May 2024	93 M MIT	0.2 GB weights, 8 GB card	not published	75.1%LIBERO	72%Octo real-robot finetuning suite (paper Table I)	not reported	Be the first
UCB Octo-Small Compact VLA	UC Berkeley	May 2024	27 M MIT	0.1 GB weights, 8 GB card	not published	not reported	not reported	83%Octo ablation suite (real WidowX)	Be the first
UMA 3D-VLA World model	UMass Amherst	March 2024	not published closed weights	not published	not published	68%RLBench	not reported	not reported	Be the first
OXE RT-1-X Action policy	Open X-Embodiment	October 2023	not published Apache-2.0	not published	not published	not reported	50%Open X-Embodiment cross-lab evaluation	not reported	Be the first
OXE RT-2-X Foundation VLA	Open X-Embodiment	October 2023	55 B closed weights	102 GB weights	not published	not reported	not reported	not reported	Be the first
RT-2 Foundation VLA	Google DeepMind	July 2023	55 B closed weights	102 GB weights	not published	90%Language-Table (simulation)	not reported	not reported	Be the first
RoboCat Action policy	Google DeepMind	June 2023	1.2 B closed weights	2.2 GB weights, 8 GB card	not published	not reported	not reported	86%RoboCat fine-tuning with 1000 demonstrations	Be the first
SU ACT (Action Chunking with Transformers) Action policy Trainable here	Stanford University	April 2023	80 M MIT	0.1 GB weights, 8 GB card	10 ms	86%ALOHA simulation (MuJoCo)	96%Real ALOHA hardware	not reported	Be the first
CU Diffusion Policy (CNN-based, DiffusionPolicy-C) Action policy	Columbia University	March 2023	278 M MIT	0.5 GB weights, 8 GB card	100 ms	93%Robomimic	not reported	not reported	Be the first
CU Diffusion Policy (Transformer-based, DiffusionPolicy-T) Action policy	Columbia University	March 2023	31 M closed weights	0.1 GB weights, 8 GB card	not published	90%Robomimic	not reported	not reported	Be the first
RT-1 Action policy	Google	December 2022	35 M Apache-2.0	0.1 GB weights, 8 GB card	not published	not reported	97%RT-1 real-robot evaluation	not reported	Be the first

Three different questions

A model that wins a simulation suite is not automatically the model that works on your bench, and neither of those tells you what happens after you fine tune it on fifty of your own episodes. The arena keeps the three apart on purpose.

Simulation

Reproducible benchmark suites. Everyone runs the same episodes, so the numbers can be compared inside a suite.

1AI2
MolmoAct2
LIBERO, Object
100%
2UNI
UnifoLM-VLA-0
LIBERO, Object
100%
3MID
DexVLA
LIBERO, Object
99.1%
4CAS
BitVLA
LIBERO, object
99%
5AI2
MolmoAct2-Think
LIBERO, Spatial
98.8%
6π
π0.5
LIBERO, Spatial, fine-tuned checkpoint at 30k steps
98.8%

Real world

Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.

1
Gemini Robotics
Gemini Robotics dexterous long-horizon specialist tasks, full long-horizon lunch-box packing
100%
2THU
RDT-1B
Real robot (Mobile ALOHA, RDT paper protocol), Pour Water-L-1/3, instruction following (correct amount poured)
100%
3GEN
GEN-1
Generalist AI internal real-robot evaluation, aggregate over the reported tasks, about 1 hour of robot data per task
99%
4
GR-2
Real robot, multi-task (ByteDance platform), 105 manipulation tasks, simple setting
97.7%
5
GR-3
Real robot (ByteMini), table bussing (long-horizon), instruction following
97.5%
6
RT-1
RT-1 real-robot evaluation, seen tasks, over 700 instructions
97%

Fine tuned tasks

What the model reaches after being adapted to a new task or a new robot, usually from a small number of demonstrations. This is the number that matters if you bring your own data.

1THU
RDT-1B
Real robot (Mobile ALOHA, RDT paper protocol), Handover, 5-shot few-shot learning
100%
2
NVIDIA Isaac GR00T N1.5
Unitree G1 post-training, Place one of two fruits or objects, 1000 demonstrations
98.8%
3AI2
MolmoAct2
Real-world zero-shot (Franka DROID setup), 5 tasks, 15 trials per task, no per-task fine-tuning
87.1%
4
RoboCat
RoboCat fine-tuning with 1000 demonstrations, KUKA lifting
86%
5π
π0.7
Physical Intelligence real-robot evaluation, Zero-shot cross-embodiment shirt folding on bimanual UR5e, no UR5e laundry data in training
85.6%
6UCB
Octo-Small
Octo ablation suite (real WidowX), 40 trials across two language-conditioned and two goal-conditioned tasks
83%

These lists rank the best value each model reports on that axis, and the suite is printed next to it. A model at the top of the simulation list is not necessarily better than the one below it if the two ran different suites. Models that also report embodied reasoning benchmarks carry those on their own page, kept separate because no robot moves in them.

What practitioners say

Benchmarks measure what the authors chose to measure. The community rating is the other half: one vote per AY-Robots account, on any model, changeable at any time. Open a model page to cast yours.

No votes yet. If you have run any of these models on real hardware, you are exactly the person whose rating is worth something here. Create an account and open any model page.

Five kinds of model

The label VLA is used loosely. These groups separate what the models actually do, which matters more for your decision than the year they came out.

Foundation VLA

39 models

Large models pretrained on multi robot corpora, meant to be fine tuned rather than trained from scratch.

Gemini Robotics 2

Google DeepMind, July 2026

not published, closed weights

X Square Robot, May 2026

4 B, Apache-2.0

GEN

GEN-1

Generalist AI, April 2026

not published, closed weights

π0.7

Physical Intelligence, April 2026

5 B, closed weights

FIG

Helix 02

Figure AI, January 2026

not published, closed weights

RBY

LingBot-VLA

Robbyant, January 2026

4 B, Apache-2.0

UNI

UnifoLM-VLA-0

Unitree, January 2026

not published, CC BY-NC-SA

SAI

InternVLA-M1

Shanghai AI Laboratory, October 2025

4.1 B, Code MIT

THU

X-VLA-0.9B

Tsinghua University, October 2025

880 M, Apache-2.0

Gemini Robotics 1.5

Google DeepMind, September 2025

not published, closed weights

THU

RDT2-FM

Tsinghua University, September 2025

not published, Apache-2.0

THU

RDT2-VQ

Tsinghua University, September 2025

8.3 B, Apache-2.0

XSR

WALL-OSS (flow)

X Square Robot, September 2025

4 B, Apache-2.0

XSR

WALL-OSS-FAST

X Square Robot, September 2025

not published, Apache-2.0

not published, closed weights

BAI

UniVLA (Unified Vision-Language-Action Model, BAAI)

BAAI, June 2025

8.5 B, Apache-2.0

NVIDIA Isaac GR00T N1.5

NVIDIA, May 2025

2.7 B, NVIDIA OneWay NC

ODL

UniVLA (task-centric latent actions)

OpenDriveLab, May 2025

7.5 B, Apache-2.0

π0.5

Physical Intelligence, April 2025

3.6 B, Apache-2.0

Gemini Robotics

Google DeepMind, March 2025

not published, closed weights

AGI

GO-1 (Genie Operator-1)

AgiBot, March 2025

3 B, CC BY-NC-SA

NVIDIA Isaac GR00T N1

NVIDIA, March 2025

2.2 B, NVIDIA OneWay NC

FIG

Helix

Figure AI, February 2025

7.1 B, closed weights

Magma-8B

Microsoft, February 2025

8.9 B, MIT

OpenVLA-OFT

Stanford University, February 2025

7.5 B, MIT

SAI

SpatialVLA

Shanghai AI Laboratory, January 2025

3.5 B, MIT

π0-FAST

Physical Intelligence, January 2025

3 B, Apache-2.0

CogACT-Base

Microsoft, November 2024

7.6 B, MIT

CogACT-Large

Microsoft, November 2024

not published, MIT

CogACT-Small

Microsoft, November 2024

not published, MIT

KST

LAPA (Latent Action Pretraining)

Tsinghua University, October 2024

1.2 B, MIT

π0 (pi-zero)

Physical Intelligence, October 2024

3.3 B, Apache-2.0

OpenVLA

Stanford University, June 2024

7.5 B, MIT

OXE

RT-2-X

Open X-Embodiment, October 2023

55 B, closed weights

RT-2

Google DeepMind, July 2023

55 B, closed weights

Compact VLA

10 models

Models small enough to fine tune on a single consumer card, and in some cases to run on the robot itself.

Gemini Robotics On-Device 2

Google DeepMind, July 2026

not published, closed weights

Chinese Academy of Sciences, June 2025

3 B, MIT

Gemini Robotics On-Device

Google DeepMind, June 2025

not published, closed weights

SmolVLA

Hugging Face, June 2025

Midea Group, February 2025

not published, closed weights

THU

RDT-170M

Tsinghua University, October 2024

170 M, MIT

MID

TinyVLA

Midea Group, September 2024

1.3 B, MIT

UCB

Octo-Small

UC Berkeley, May 2024

27 M, MIT

Reasoning VLA

17 models

Models that emit an intermediate plan, a subtask or a chain of thought before they emit actions.

ACE

ACE-Brain-0.5

ACE-Brain, July 2026

8.8 B, see model card

Gemini Robotics ER 2

Google DeepMind, July 2026

not published, closed weights

AI2

MolmoAct2

Allen Institute for AI, May 2026

5.5 B, Apache-2.0

AI2

MolmoAct2-Think

Allen Institute for AI, May 2026

5.5 B, Apache-2.0

Gemini Robotics-ER 1.6

Google DeepMind, April 2026

not published, closed weights

AGI

GO-2 (Genie Operator-2)

AgiBot, April 2026

not published, closed weights

NVIDIA Isaac GR00T N1.7

NVIDIA, April 2026

3.1 B, NVIDIA Open Model

NVIDIA Isaac GR00T N1.6

NVIDIA, December 2025

3.3 B, NVIDIA OneWay NC

π*0.6

Physical Intelligence, November 2025

not published, closed weights

Gemini Robotics-ER 1.5

Google DeepMind, September 2025

not published, closed weights

AI2

MolmoAct-7B-D

Allen Institute for AI, August 2025

8.1 B, Apache-2.0

AI2

MolmoAct-7B-O

Allen Institute for AI, August 2025

7.7 B, Apache-2.0

MID

ChatVLA-2

Midea Group, May 2025

not published, closed weights

Gemini Robotics-ER

Google DeepMind, March 2025

not published, closed weights

MID

ChatVLA

Midea Group, February 2025

3.4 B, MIT

UCB

Embodied Chain-of-Thought (ECoT)

UC Berkeley, July 2024

7.2 B, MIT

PKU

RoboMamba

Peking University, June 2024

3.2 B, closed weights

Action policy

15 models

Policies without language conditioning. They learn one task from your demonstrations and are the baseline every VLA has to beat.

Shanghai AI Laboratory, December 2024

316 M, closed weights

SAI

Seer-Large

Shanghai AI Laboratory, December 2024

566 M, closed weights

UC Berkeley, August 2024

130 M, MIT

UCB

Octo-Base

UC Berkeley, May 2024

93 M, MIT

OXE

RT-1-X

Open X-Embodiment, October 2023

not published, Apache-2.0

RoboCat

Google DeepMind, June 2023

1.2 B, closed weights

ACT (Action Chunking with Transformers)

Stanford University, April 2023

80 M, MIT

Diffusion Policy (CNN-based, DiffusionPolicy-C)

Columbia University, March 2023

278 M, MIT

Diffusion Policy (Transformer-based, DiffusionPolicy-T)

Columbia University, March 2023

31 M, closed weights

RT-1

Google, December 2022

35 M, Apache-2.0

World model

4 models

Models that predict future observations. They are used to generate training data or to plan, not to command joints directly.

Alibaba, November 2025

not published, Apache-2.0

GR-2

ByteDance, October 2024

230 M, closed weights

UMA

3D-VLA

UMass Amherst, March 2024

not published, closed weights

5 of these run on your own arm from here

Record a dataset with an SO-100 class arm, pick a model, and the training pool rents the GPU by the hour. For those models the platform pages carry measured numbers from that pool rather than published ones: real defaults, real GPU tier, real inference latency.

NVIDIA Isaac GR00T N1.7 SmolVLA NVIDIA Isaac GR00T N1.5 ππ0.5 SUACT (Action Chunking with Transformers)

Platform model pages Training guides Get the client

SO-100 robot arm, the reference hardware of the AY-Robots platform — An SO-100, the reference arm of this platform. Roughly 110 to 150 euro in parts, six Feetech STS3215 servos at 7.4 V.

How this list is built

A comparison table is only worth reading if you know where its numbers came from. Four rules, applied to every row.

Every number carries its source

Parameter counts, licences, latencies and benchmark results are taken from the paper, the model card or the official repository of the model in question, and the link sits next to the value. Nothing here is inferred from a similar model.

Missing is missing

Most publications do not report inference latency or GPU memory. Those fields stay empty instead of being filled with a plausible guess. An empty cell is a correct answer, an invented one is not.

Computed values are labelled

The only numbers the arena produces itself are the weights only memory footprint at a given precision and the smallest real GPU that fits it. Both are derived from the parameter count with a stated formula, and both are labelled as computed wherever they appear.

Comparison stays inside a suite

Results are grouped into simulation, real world and fine tuned tasks, and each value keeps its suite and split. Two models are only directly comparable when they report the same suite on the same split.

Last editorial update 2026-08-11. Company names and logos are the trademarks of their owners and appear here only to identify who published which model.

Questions

What is a vision language action model?: A vision language action model, usually shortened to VLA, takes camera images and a task written in plain language and outputs robot actions, normally joint positions or end effector deltas. The vision and language part is typically a pretrained vision language model, and an action head converts its representation into continuous commands. The point of the design is that a single policy can follow instructions it was never explicitly trained on, instead of one policy per task.
Why does the arena not rank models with one overall score?: Because the numbers come from different benchmark suites, and a LIBERO success rate and a real robot success rate do not measure the same thing. Averaging them would produce a number that looks authoritative and means nothing. The arena therefore sorts inside a column, always shows which suite a value came from, and links every value to the source it was taken from.
How much GPU memory does a VLA need?: Most papers do not publish a memory figure, so the arena computes the weights only footprint from the parameter count at bf16 precision and marks it as computed. Weights are a lower bound: activations, the image encoder intermediates, the KV cache and the CUDA context come on top, which in practice is roughly 30 to 50 percent more for inference. Fine tuning is a different order of magnitude again, because optimiser state and gradients usually cost several times the weights unless you use LoRA or a similar adapter.
Which of these models can I actually train on AY-Robots?: Five of them: GR00T N1.7, GR00T N1.5, Pi0.5, SmolVLA and ACT. Those are the models the training pool can fine tune on a LeRobot dataset that you record with your own arm, and the platform pages for them carry the real defaults, GPU tier and measured inference latency from that pool. Every other model in the arena is listed for comparison and links to its own repository.
What does the community rating mean?: It is a one to five rating from people with an AY-Robots account, one vote per account and per model, changeable at any time. It is a subjective signal about how a model behaved in practice, not a benchmark. Read it next to the measured numbers, not instead of them.
How current is this list?: Every entry carries its own editorial date and the sources it was built from. The field moves fast, so a model can gain a new checkpoint or a new licence between updates. Where a source contradicts what you read here, trust the source and tell us, the link is on every model page.

Drive a real arm before you pick a model

A physical SO-100 is online and free to control in the browser. No signup, no client, no hardware of your own.

Download the client Drive a real arm now