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.

THU

X-VLA-0.9B

Tsinghua University, China · October 2025

Institute for AI Industry Research (AIR) Tsinghua University, Shanghai Artificial Intelligence Laboratory, Peking University

Foundation VLAOpen weights, Apache-2.0Also written X-VLA, X-VLA-Pt, lerobot/xvla-base

Parameters: 880 M; The released 2toINF/X-VLA-Pt checkpoint reports 879,738,545 F32 parameters in its safetensors index (Hugging Face model API,...
GPU memory: 1.6 GB; weights at bf16, computed. Fits an 8 GB card
Inference latency: not published; per action step
Weights: Apache-2.0; downloadable checkpoint

What it is

X-VLA is a cross-embodiment VLA that handles heterogeneous robot data with soft prompts, a small set of learnable embeddings per data source that are injected from the earliest stage of multimodal fusion. The architecture is deliberately plain: standard Transformer encoders (hidden size 1024, 24 blocks) over a Florence-2-Large vision language encoder, with a flow matching action head and 10 denoising steps, totalling 0.88B parameters. Phase I pretrains on 290K episodes from DROID, RoboMIND and AgiBot spanning seven platforms and five arm types; Phase II adapts to a new robot by learning a fresh set of soft prompts. The 0.9B instantiation sets state of the art across LIBERO (98.1 average), SimplerEnv WidowX (95.8), CALVIN ABC to D (4.43), RoboTwin 2.0 and VLABench while being three to eight times smaller than the 3B and 7B baselines it beats. Accepted at ICLR 2026, Apache-2.0, and natively integrated into LeRobot. The gap in the record is deployment data: the paper publishes no latency, no throughput and no VRAM figure on any named GPU.

Architecture

Backbone: Florence-2-Large vision language encoder (microsoft/Florence-2-large), with a shared vision backbone for auxiliary views such as wrist cameras
Action head: Flow matching over standard Transformer encoders. LeRobot configuration: hidden_size 1024, depth 24, num_heads 16, num_denoising_steps 10, len_soft_prompts 32, num_domains 30. Unified EE-6D control space: Cartesian end-effector xyz, absolute rotation in Rotate6D representation, and discretized binary gripper state, padded to 20 dimensions.
Action chunk: 32 steps per forward pass
Parameters: The released 2toINF/X-VLA-Pt checkpoint reports 879,738,545 F32 parameters in its safetensors index (Hugging Face model API, checked 11.08.2026), consistent with the paper's '0.9B instantiation'. Soft prompts account for 0.04 percent of total parameters. PEFT adaptation to a new embodiment tunes 9M parameters, about 1 percent, which is 300 times fewer trainable parameters than the 3B pi-0 baseline.
Pretraining data: Phase I: 290K episodes from DROID, RoboMIND and AgiBot, spanning seven platforms across five types of robotic arms from single-arm to bi-manual. Composition by share: AGIBOT 48.8 percent at 30 Hz (head and wrist cameras), Franka 15.8 percent at 15 Hz from DROID (left/wrist and right/wrist), Franka 6.7 percent at 30 Hz from RoboMIND (top), UR-5 8.7 percent at 30 Hz, Agilex 3.7 percent at 30 Hz, Dual-Franka 0.8 percent at 30 Hz. The Hz values describe the source datasets, not the policy control rate.
Embodiments: Franka Panda (LIBERO, CALVIN, VLABench), WidowX (BridgeData V2, SimplerEnv), Google Robot (SimplerEnv), AgiBot G1 (AgiBot World Challenge, IROS 2025 winner), Agilex (RoboTwin 2.0, SoftFold cloth folding), AIRBOT, SO-101 bimanual (supported action mode in LeRobot)

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 880 M 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	3.3 GB	Training master weights, rarely used for inference
bf16 or fp16	1.6 GB	The usual way these checkpoints are served
int8	0.8 GB	Quantised, expect some loss of precision on fine motions
int4	0.4 GB	Aggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 8 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 minimum GPU or VRAM figure is published. The LeRobot documentation only advises setting policy.dtype=bfloat16 to avoid OOM during fine-tuning, and lists reducing chunk_size from 32 to 16, enabling gradient checkpointing, reducing batch size or freezing more components as the OOM remedies.

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 4 suites
98.1%
success ratePaper main results table. Per split: Spatial 98.2, Object 98.6, Goal 97.8, Long 97.6. Best prior SOTA in the same table is 97.1 (OpenVLA-OFT, 7B).source
SimplerEnv (WidowX / Bridge) average
95.8%
success ratePaper main results table; best prior SOTA in that table is 71.9 (MemoryVLA). Treat with caution: the ACE-Brain-0.5 paper lists X-VLA at 76.0 average on SimplerEnv-Bridge under a 120-step inference horizon, quoting the GTA-VLA paper. Evaluation protocols likely differ.source
SimplerEnv (Google Robot) Visual Matching
80.4%
success ratePaper main results table. The GitHub model zoo reports 83.5 percent VM for the released X-VLA-Google-Robot checkpoint, so the shipped checkpoint differs slightly from the paper.source
RoboTwin 2.0 Easy
70%
success ratePaper main results table; best prior SOTA 46.4 (pi-0). Hard split 39.0 versus prior 16.4. Trained with 50 demonstrations per task on the Agilex dual-arm platform.source
VLABench average
51.1 score
progress scorePaper main results table; best prior SOTA 39.7 (GR00T-N1).source
CALVIN ABC to D
4.43 tasks
average successful sequence length (max 5)Paper main results table. Below FLOWER at 4.53, so X-VLA is not top on CALVIN.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

Evaluated on three physical robot platforms. Cloth folding on Agilex reaches 100 percent success over 2 hours of continuous operation after adaptation on the Soft-FOLD dataset. Won first place at the AgiBot World Challenge at IROS 2025 on the AgiBot G1. Real-world pick-and-place follows the BridgeData V2 protocol on WidowX. Per-task real robot success rates are reported qualitatively on the project page rather than as a numeric table.

Fine tuning it yourself

Two paths. Full fine-tune is the official recipe and gives best results: do not freeze the VLM encoders, train the policy transformer and the soft prompts, and train the VLM at one tenth of the base learning rate (LeRobot applies this LR ratio by default). PEFT is the cheap path: tune only the 9M soft prompt parameters, roughly 1 percent, which reaches 93 percent on LIBERO and 54.2 percent on Simpler-WidowX. Data must be in LeRobot dataset format; action_mode=auto detects the dataset action dimension and pads to the model's 20 dimensions. Pick a domain_id matching your setup (LIBERO uses 3, Bridge 0, RT1 1, CALVIN 2). Cloth folding was learned from a 2-hour dataset.

Where it helps, where it does not

Strengths

State of the art across six simulation suites at 0.9B parameters, beating 3B and 7B models including pi-0, OpenVLA-OFT, GR00T-N1 and UniVLA
Adaptation to an unseen embodiment needs only 9M trainable parameters, 300 times fewer than the pi-0 baseline, while staying near pi-0 quality
Apache-2.0 with weights, training code and evaluation harnesses per benchmark, plus native LeRobot integration
Best LIBERO-Long score in its own comparison table (97.6), where most models lose the most ground
Scaling curves show no saturation at 0.9B parameters and 290K episodes, so the recipe has headroom
Validated on real hardware including an IROS 2025 challenge win, not simulation only

Limits

No latency, throughput or VRAM measurement on any named GPU is published in the paper, the repository or the LeRobot documentation
SimplerEnv Google Robot numbers differ between artefacts: the paper reports 80.4 VM and 75.7 VA while the GitHub model zoo lists 83.5 VM and 76.4 VA for the released checkpoint
The 95.8 percent SimplerEnv WidowX figure is far above what third parties reproduce; the ACE-Brain-0.5 paper lists X-VLA at 76.0 average on SimplerEnv-Bridge, quoted from the GTA-VLA paper
PEFT mode costs real accuracy on long horizon tasks: LIBERO-Long drops from 97.6 full fine-tune to 84.2 with 9M tuned parameters
Domain ID handling is a manual step and a documented source of low success rates when set incorrectly
Model output is fixed at 20 action dimensions and requires padding and trimming for robots with different action spaces
Matching the officially reported numbers may require an additional soft-prompt warm-up LR schedule that is not implemented in LeRobot

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