| --- |
| library_name: servovla |
| license: mit |
| tags: |
| - robotics |
| - vision-language-action |
| - lerobot |
| - libero |
| pipeline_tag: robotics |
| --- |
| |
| # ServoVLA LIBERO |
|
|
| ServoVLA policy for LIBERO-40 simulation. This release contains the EMA policy |
| head from training step 1,190,000 and the action normalization statistics saved |
| with that run. |
|
|
| ## Model Contract |
|
|
| - Benchmark: LIBERO-40 (`libero_10`, `libero_spatial`, `libero_object`, `libero_goal`) |
| - Action space: 7-dimensional direct environment action |
| - State space: 8-dimensional LIBERO proprioception |
| - Cameras: `observation.images.image`, `observation.images.image2` |
| - Action chunk: 16 steps |
| - Inference steps: 3 |
| - Vision encoder: `facebook/dinov3-vitb16-pretrain-lvd1689m` |
| - Language encoder: `Qwen/Qwen3.5-0.8B` |
| - Weights: EMA (`used_ema=true`) |
| - Action normalization: enabled, per-horizon dimension, 16 x 7 statistics |
|
|
| ## Evaluation |
|
|
| The selected checkpoint was evaluated on the fixed LIBERO-40 four-suite |
| protocol with 30 episodes per task, 1,200 episodes total: |
|
|
| - Overall balanced success rate: 78.0% |
| - `libero_10`: 79.0% |
| - `libero_spatial`: 75.0% |
| - `libero_object`: 84.33% |
| - `libero_goal`: 73.67% |
|
|
| Three additional 50-episode-per-task evaluations with different seeds produced |
| overall balanced success rates of 75.6%, 79.15%, and 69.25% (mean 75.95%). |
| Rollouts used relative LIBERO environment control, open-loop 16-step chunks, |
| and observed maximum frame delay 0 in the reported runs. |
|
|
| ## Files |
|
|
| - `model.safetensors`: LeRobot-compatible EMA policy weights. |
| - `config.json`: ServoVLA policy configuration. |
| - `policy_preprocessor.json` and `policy_postprocessor.json`: LeRobot processor configuration. |
| - `checkpoint.pt`: cleaned ServoVLA checkpoint containing `model_state`, `ema_state`, |
| `step`, and `action_normalization` only. |
|
|
| The checkpoint is weights-only for inference and fresh initialization. Optimizer, |
| scheduler, W&B, and resume-training state are intentionally absent. |
|
|
| ## Usage |
|
|
| Use this directory with the ServoVLA LIBERO evaluation entry points. The LIBERO |
| camera keys, state ordering, action dimensions, and direct-action semantics must |
| match the model contract above. This policy is not compatible with SO101 robot |
| observations or the SO101 delta-action contract. |
|
|
| ## License |
|
|
| ServoVLA is released under the MIT License. The vision and language encoders and |
| the LIBERO benchmark assets are subject to their respective upstream licenses |
| and terms. |
|
|
