RoboSynChallenge ACT β€” handle_basket

Fine-tuned ACT policy checkpoint for the RoboSynChallenge handle_basket task.

Task

Pick up the bottle, place it into the basket, then grasp and carry the basket with the other arm.

Model

  • Policy: ACT (Action Chunking with Transformers)
  • Framework: RoboSynChallenge / LeRobot
  • Input: three RGB camera views and robot state
  • Output: dual-arm joint and gripper actions
  • Checkpoint format: LeRobot pretrained_model
  • Initialization: official RoboSynChallenge ACT checkpoint followed by task-specific fine-tuning

The complete checkpoint directory is provided. Keep model.safetensors, config.json, and train_config.json together when downloading or evaluating the model.

Evaluation results

  • Standard random protocol: 138 successes in 250 episodes (55.2%).
  • Strict paired-scene grip-balanced comparison: 32/50 (64.0%), compared with 15/50 (30.0%) for the official ACT checkpoint on the same 50 scenes.

These figures were measured with the official success condition. See the associated GitHub submission for exact seeds, pinned checkpoint revision, and evaluation commands.

Download

hf download RfZnCu/robosyn-act-handle-basket \
  --local-dir checkpoints/handle_basket

For reproducible evaluation, pin the full Hugging Face commit revision recorded in the associated GitHub submission.

Intended use

This checkpoint is intended for RoboSynChallenge research and evaluation. It is not a general real-world robot safety controller. Validate actions in simulation before deploying on hardware.

Attribution and terms

This is a fine-tuned checkpoint derived from the official RoboSynChallenge ACT checkpoint and datasets. Usage remains subject to the original RoboSynChallenge project, checkpoint, and dataset terms. No separate model-weight license is asserted by this model card.

Downloads last month
7
Safetensors
Model size
51.6M params
Tensor type
F32
Β·
Video Preview
loading