ted88168/rollout_colorlogo_rl_round2_combined

SO-101 Pi0.5 ACP Round 2 Combined Training Pool

This LeRobot dataset is the immutable training pool for the second round of the Pi0.5 Advantage-Conditioned Policy (ACP) workflow.

Dataset summary

• Episodes: 204
• Frames: 239,093
• FPS: 30
• Cameras: observation.images.front and observation.images.handeye
• Robot state/action dimensions: 6
• Successful episodes: 127
• Failed episodes: 77
• Round 1 contribution: 120 episodes and 137,200 frames
• Round 2 contribution: 84 episodes and 101,893 frames

Round 1 contains the original complete human-in-the-loop training trajectories. Round 2 contains targeted autonomous rollouts selected from failure conditions found during fixed-matrix hardware evaluation. Source rounds remain available as separate immutable datasets.

Pi0.5 ACP enhancement

This project extends Pi0.5 training with an experimental reinforcement-learning data loop:

1. train a trajectory value model from complete successful and failed episodes;
2. infer per-frame value and n-step advantage;
3. convert advantage into a binary ACP conditioning label;
4. fine-tune Pi0.5 with advantage-conditioned prompts and indicator dropout;
5. evaluate against a frozen baseline and collect new complete episodes from failed conditions.

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

Features used by the ACP loop

The aggregate preserves observations, actions, task text, timestamps, episode indices, and episode-level success labels. Round-source and intervention-related auxiliary fields are retained where available. Failed episodes are intentionally included for value learning.

Validation

The aggregate passed the following checks before upload:

• all 204 episode indices are present and unique;
• aggregate frame and outcome counts equal the source manifests;
• success labels are constant within each episode;
• both camera streams are present;
• representative Round 1 H.264 and Round 2 AV1 episodes decode through the LeRobot PyAV reader;
• source videos were copied without re-encoding;
• source datasets were not overwritten.

The aggregate contains separate H.264 and AV1 MP4 files. Consumers should use a video backend that probes each file directly.

Intended use

The dataset is intended for research on value learning, offline advantage inference, advantage-conditioned behavior cloning, and reproducible comparison of sequential robot-policy improvement rounds.

It is not a general-purpose robotics benchmark and should not be treated as evidence of safety or performance outside the recorded SO-101 setup.

Related resources

• Source code: BurningDawn8888/lerobot-pi05-acp
• Base framework: Hugging Face LeRobot
• Base policy family: Pi0.5
• Round 1 dataset: ted88168/rollout_colorlogo_rl_round1
• Targeted Round 2 source: ted88168/rollout_colorlogo_acp_r2_targeted_v1
• Fixed-matrix evaluation: ted88168/rollout_colorlogo_stage6_eval

Safety and limitations

Real-robot policies trained from this dataset can move hardware unexpectedly. Validate calibration, camera mappings, action ranges, reset poses, and emergency-stop access before deployment. The data reflects one robot, workspace, lighting setup, camera arrangement, object family, and task distribution.