Asset Description

This directory contains the small-scale tutorial images used in the AMD ROCm policy replication project. The charts and keyframe sequences are generated by ../code/generate_tutorial_assets.py to verify the success rate and rollout behavior in the tutorials; the screenshots in aup_cloud_guide/ come from the AMD / AUP cloud platform usage guide compressed package.

When generating the chart again, first prepare your experiment output directory, and then run:

bashpython code/generate_tutorial_assets.py --source-root "$OUTPUT_ROOT"

$OUTPUT_ROOT should include batch evaluation JSONL/TSV and representative rollout videos. If --source-root is not provided, the script will generate charts using built-in teaching metrics and an indicator summary chart for missing videos with source explanations.

File	Purpose
`model_status_summary.png`	Overview of the current reimplementation status for the topic
`smolvla_red_blue_success.png`	Comparison of the success rate of fixed instructions for SmolVLA red/clear cups
`act_dagger_progress_curve.png`	Change in success rate of ACT from baseline to DAgger correction
`smolvla_blue_failure_sequence.jpg`	Critical frames for failure of SmolVLA baseline clear cups
`smolvla_blue_success_sequence.jpg`	Critical frames for success of clear cups after SmolVLA weighted sampling
`act_failure_sequence.jpg`	Typical critical frames for failure of ACT DAgger
`act_success_sequence.jpg`	Critical frames for physical success of ACT DAgger
`pi0_raw_vs_finisher_diagnostic.png`	Diagram of diagnostic metrics at the end of pi_0 raw and script terminator
`pi0_ep2_raw_vs_finisher_side_by_side.mp4`	Comparison video of pi_0 episode2 raw-vs-hybrid
`pi0_ep2_raw_vs_finisher_frame.png`	Critical frames in the comparison video of pi_0 episode2
`pi0_ep2_raw_vs_finisher_metrics.md`	Summary table of metrics for pi_0 episode2 raw-vs-hybrid
`metrics_snapshot.json`	Snapshot of small-volume metrics, including diagnosis for ACT, SmolVLA, pi_0 small set, pi_0 full20 open-loop, and closed-loop strict numbers
`collection_dataset_snapshot.json`	Summary of measurements on a AMD device for 20 episodes, 2621 frames, and 20 Hz dataset
`training_progress_snapshot.json`	Summary of historical training nodes for ACT, SmolVLA, pi0, and learned head
`training_progress_overview.png`	Overview of historical training steps and closed-loop results, to avoid misinterpreting training completion rate as success rate
`pi0_strict_input_results.json`	Final strict results for raw/head, fixed/random environments, and seed bug
`pi0_strict_input_progress.png`	Comparison diagram of fixed scenarios and random environments for pi0 strict-input
`pnp_four_view_strict_success.mp4`	Video of strict success measurements from four perspectives: Agent/Egocentric/Top/Side
`pnp_four_view_strict_success_sequence.jpg`	Sequence of critical frames at 5 moments in the four-perspective video
`aup_cloud_guide/`	Screenshots of the AUP Learning Cloud JupyterHub / Code Server usage guide (primary entry)
`amd_radeon_cloud/`	Official tutorial images for AMD Radeon Cloud Developer Cloud and screenshots of the AMD ROCm Embodied AI Policy Replication workspace (alternative entry)