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.

BAI

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

BAAI, China · June 2025

Beijing Academy of Artificial Intelligence (BAAI) with CASIA

Foundation VLAOpen weights, Apache-2.0Also written BAAI UniVLA, Unified VLA

Parameters: 8.5 B; The paper states a purely autoregressive transformer with 8.5 billion parameters, identical in architecture to Emu3.
GPU memory: 16 GB; weights at bf16, computed. Fits a 24 GB card
Inference latency: not published; per action step
Weights: Apache-2.0; downloadable checkpoint

What it is

This UniVLA, from BAAI with CASIA, is a different model from the OpenDriveLab UniVLA despite the identical name, and the two are frequently confused. It is an 8.5B parameter autoregressive transformer with the Emu3 architecture that treats vision, language and action as one interleaved sequence of discrete tokens, with images vector-quantized and actions discretized through a Discrete Cosine Transform into the frequency domain. Post-training adds world modeling from video so the model learns causal dynamics before policy learning, which the authors argue is what carries long-horizon tasks. It posts the strongest simulation numbers in this round: 95.5 percent LIBERO average, 4.41 CALVIN ABC-D average length and 69.8 percent on SimplerEnv-Bridge against 42.7 percent for the prior best. The two things a practitioner needs are missing: there are no real-robot results at all, and no licence is stated on the repository or the model page.

Architecture

Backbone: Emu3 (Emu3-base and Emu3-vision pretraining checkpoints from BAAI)
Action head: No separate action head. Vision, language and action are all discretized into one interleaved token sequence and modelled autoregressively. Images are discretized with a vector-quantized encoder; actions are transformed into the frequency domain and discretized via Discrete Cosine Transform (DCT) encoding.
Parameters: The paper states a purely autoregressive transformer with 8.5 billion parameters, identical in architecture to Emu3. This is the largest model in this comparison round.
Pretraining data: Built on Emu3-base and Emu3-vision. Post-training adds world modeling from video to capture causal dynamics before downstream policy learning. Separate image-SFT and video-SFT variants are released per benchmark.
Embodiments: CALVIN simulation (ABC-D and ABCD-D), LIBERO simulation, SimplerEnv WidowX / Bridge

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

Computed from 8.5 B parameters. Weights only, so treat it as a floor: activations, image encoder intermediates and the CUDA context come on top, in practice 30 to 50 percent more for inference.
Precision	Weights	Note
fp32	32 GB	Training master weights, rarely used for inference
bf16 or fp16	16 GB	The usual way these checkpoints are served
int8	7.9 GB	Quantised, expect some loss of precision on fine motions
int4	4.0 GB	Aggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 24 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.

No inference VRAM is published. Post-training ran on 32 A100 40GB GPUs for roughly 4 to 5 days over 50k steps, which is the only hardware datapoint in the paper. At 8.5B parameters this is the heaviest model here.

The authors publish no memory or latency figure for this model. Everything above is computed from the parameter count.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

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

LIBERO average over Spatial, Object, Goal and Long, video SFT
95.5%
success rateAuthors' own evaluation. Per-suite: Spatial 95.4, Object 98.8, Goal 93.6, Long 94.0. The image-SFT variant averages 94.8. pi-0-FAST is cited at 85.5 percent.source
SimplerEnv (WidowX / Bridge) average over four tasks
69.8%
success rateAuthors' own evaluation. Per-task: Put Spoon 83.3, Put Carrot 66.7, Stack Block 33.3, Put Eggplant 95.8. Prior state of the art cited at 42.7 percent. Stack Block remains the weakest task, as with the OpenDriveLab UniVLA.source
CALVIN ABCD-D, video SFT
4.63 tasks (out of 5)
average length (tasks completed in sequence)Authors' own evaluation. Rises to 4.71 with 5x inference steps (180 total), which quantifies the accuracy-versus-compute tradeoff.source
CALVIN ABC-D (zero-shot environment D)
4.41 tasks (out of 5)
average length (tasks completed in sequence)Authors' own evaluation. ABC-D is the harder split because environment D is unseen during training.source

Real world

No results in this category are published for this model.

Fine tuned tasks

No results in this category are published for this model.

On real hardware

No real-robot results are reported. All published evidence is in simulation (CALVIN, LIBERO, SimplerEnv-Bridge), which is the single biggest gap in this entry relative to the other models in the comparison.

Fine tuning it yourself

Released checkpoints are per-benchmark: a world-model post-training checkpoint plus separate CALVIN, LIBERO (image-SFT and video-SFT) and SimplerEnv-Bridge fine-tunes. Post-training cost was 32 A100 40GB GPUs for 4 to 5 days. No licence is stated on the repository or the model page, which is a real blocker for commercial use until clarified.

Where it helps, where it does not

Strengths

Best SimplerEnv-Bridge result in this comparison at 69.8 percent average, against 42.7 percent for the prior state of the art and 47.9 percent for the similarly named OpenDriveLab UniVLA
One architecture handles perception grounding, world modeling and policy learning with no task-specific heads, since everything is a discrete token sequence
Strong on both CALVIN splits: 4.41 average length on ABC-D (the harder zero-shot-environment split) and 4.63 on ABCD-D
LIBERO average of 95.5 percent with video-SFT
Inference cost is explicitly tunable: 5x inference steps raises CALVIN ABCD-D from 4.63 to 4.71
Accepted at ICLR 2026

Limits

No real-robot results at all; every published number is simulation, which makes it the weakest-evidenced model here for a hardware decision
No licence is stated on the GitHub repository or the Hugging Face model page, so commercial use is legally unclear
8.5B parameters is the largest in this comparison, with no published inference VRAM or latency
DCT-based action discretization means the action representation is frequency-domain, which is unusual and may complicate integration with existing control stacks
Released checkpoints are per-benchmark specialists rather than one generalist model
Name collision with the unrelated OpenDriveLab UniVLA published one month earlier, so cite by arXiv ID (2506.19850) to avoid mixing up the two

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-08-11.

Try a policy on a real arm

A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.

Download the client Drive a real arm now