Robotics
LeRobot
Safetensors
pistar06
pi05
advantage-conditioned-policy

Colorlogo Pi0.5 ACP Round 1 Value Model

This is the trajectory-value model used in the first Pi0.5 Advantage-Conditioned Policy (ACP) training round for an SO-101 color-block manipulation task.

It is not a robot-control policy and should not be passed directly to lerobot-rollout. It estimates task-progress value for offline value, n-step advantage, and binary ACP-label inference.

Role in Pi0.5 ACP

The experimental ACP loop is:

  1. collect complete successful, intervention-recovery, and failed episodes;
  2. train this trajectory-value model from episode outcomes;
  3. infer per-frame value and n-step advantage;
  4. produce a binary ACP conditioning label;
  5. fine-tune Pi0.5 using advantage-conditioned prompts and indicator dropout;
  6. compare the ACP checkpoint with a frozen baseline on the same real-robot matrix.

The implementation and command templates are open source at BurningDawn8888/lerobot-pi05-acp. This is an experimental extension built on Hugging Face LeRobot and is not an official Pi0.5 feature.

Training data

  • Dataset: ted88168/rollout_colorlogo_rl_round1
  • Episodes: 120
  • Frames: 137,200
  • Outcome composition: 51 autonomous successes, 33 successful intervention recoveries, and 36 failures
  • Cameras: front and hand-eye

Training configuration

Setting Value
Training steps 8,000
Batch size 8
Optimizer AdamW
Learning rate 5e-5
Precision bfloat16
LeRobot version 0.6.2 compatibility layer

Intended use

Use this checkpoint with the repository's lerobot-value-infer workflow to produce value, advantage, and ACP indicator fields on a compatible dataset copy. Validate field coverage, binary indicators, positive ratios, source immutability, and held-out trajectories before policy fine-tuning.

Limitations

The model was trained on one SO-101 setup and task family. It may learn correlations with camera viewpoint, lighting, background, intervention behavior, or task distribution. It must be validated on held-out episodes and must not be treated as a safety controller.

Related resources

Safety

This value model does not make real-robot deployment safe. Any downstream policy requires independent calibration, action-range, reset-pose, camera-mapping, and emergency-stop validation.

Downloads last month
9
Safetensors
Model size
1B params
Tensor type
F32
·
BF16
·
Video Preview
loading

Dataset used to train ted88168/colorlogo_value_round1