ted88168/rollout_colorlogo_rl_round1

SO-101 Pi0.5 ACP Round 1 Rollouts

This dataset contains 120 complete SO-101 episodes used for the first Pi0.5 Advantage-Conditioned Policy (ACP) training round.

Summary

• Episodes: 120
• Frames: 137,200
• Cameras: front and hand-eye
• Autonomous successes: 51
• Successful intervention recoveries: 33
• Failures: 36

Episodes are recorded from the initial state through the terminal outcome. Observations, policy actions, executed actions, intervention state, task text, and episode-level outcomes are retained for value learning and ACP label generation.

Method

Pi0.5 ACP adds a trajectory-value and offline-advantage loop to Pi0.5 training. A value model learns from successful and failed complete episodes. Per-frame n-step advantages are converted to a binary conditioning signal used during Pi0.5 fine-tuning with indicator dropout.

Source code and reproducible workflow: BurningDawn8888/lerobot-pi05-acp.

This is an experimental extension built on Hugging Face LeRobot, not an official Pi0.5 feature.

Related resources

• Value model: ted88168/colorlogo_value_round1
• Fixed-matrix evaluation: ted88168/rollout_colorlogo_stage6_eval
• Round 2 training pool: ted88168/rollout_colorlogo_rl_round2_combined

Limitations

The dataset represents one SO-101 setup, workspace, lighting arrangement, camera configuration, object family, and task distribution. It is intended for robotics research and does not establish deployment safety.