ServoVLA/ServoVLA-SO101

ServoVLA SO101

ServoVLA policy for SO101 real-robot manipulation. This release contains the EMA policy head from training step 150,000 and the action normalization statistics computed from the three public ServoVLA training datasets.

Model Contract

• Action space: 6-dimensional delta action
• State space: 6-dimensional SO101 proprioception
• Cameras: observation.images.front, observation.images.wrist, observation.images.side
• 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 6 statistics

Training Data

The normalization statistics were computed with equal dataset weighting over:

• ServoVLA/so101_clean_train
• ServoVLA/so101_pick_place_train
• ServoVLA/so101_stacking_train

Validation datasets were not used to compute normalization statistics.

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 codebase and its LeRobot policy loader. The SO101 camera keys and state ordering must match the model contract above. The policy is not compatible with LIBERO observation or action dimensions.

License

ServoVLA is released under the MIT License. The vision and language encoders are distributed by their respective upstream projects and remain subject to their own licenses and terms.