ctegdf/pi05-libero-rollouts

pi0.5 / OpenVLA / ACT LIBERO evaluation rollouts

Evaluation code, results, and reports: GitHub repository

Rendered rollout videos for three policies evaluated under the same LIBERO / LIBERO-Plus / LIBERO-Pro protocol: π0.5 (pi05_libero, single unified checkpoint), OpenVLA-7B, and ACT (both OpenVLA and ACT use 4 independently fine-tuned/trained per-suite checkpoints, not one unified weight). The companion GitHub repository contains the evaluation harness, summaries, and episode metadata; OpenVLA/ACT results additionally follow the versioned MODEL_EVAL_RESULT_PROTOCOL_V1 layout under results/runs/<model_id>/<benchmark_id>/<run_id>/.

Coverage

π0.5 (videos/{official,base-libero-assets,plus,pro,libero-x}/):

• LIBERO official and smoke controls
• LIBERO-Plus full and smoke runs
• LIBERO-Pro full and smoke runs (the available 8,000-episode matrix)
• LIBERO-X LEVEL1-5 full and smoke runs

OpenVLA / ACT (videos/{openvla,act}/<benchmark>/<run_id>/, protocol v1 layout):

• LIBERO (libero), LIBERO-Plus (libero-plus), LIBERO-Pro (libero-pro) — full phase only, 4 per-suite checkpoints each

Source	Benchmark	Full	Smoke	MP4s
π0.5	LIBERO official	2,000	8	2,008
π0.5	LIBERO Base assets	2,000	8	2,008
π0.5	LIBERO-Plus	10,030	28	10,058
π0.5	LIBERO-Pro	8,000	16	8,016
π0.5	LIBERO-X	34,520	29	34,549
OpenVLA	LIBERO	2,000	—	2,000
OpenVLA	LIBERO-Plus	10,030	—	10,030
OpenVLA	LIBERO-Pro	8,000	—	8,000
ACT	LIBERO	2,000	—	2,000
ACT	LIBERO-Plus	10,030	—	10,030
ACT	LIBERO-Pro	8,000	—	8,000
Total				96,699

manifests/video-manifest.jsonl (π0.5) and manifests/video-manifest.jsonl (shared, appended by the OpenVLA/ACT uploader) record the byte size and SHA-256 for every file.

All 6,000 LIBERO-X LEVEL1 episode records resolve to non-empty source MP4s in the current archive. Eight LIBERO-Plus smoke records (π0.5 only) are explicitly marked video_available=false because their run status is not_recorded. Every OpenVLA/ACT full-phase episode has a matched video (videos_available == total for all 6 runs, per each run's run.json). Missing records are retained in companion metadata rather than represented by fabricated files.

Interpretation

LIBERO-Pro has 2,000 unavailable env episodes per policy; they are N/A, not failures. LIBERO-X (π0.5 only) is a zero-shot transfer evaluation of a checkpoint fine-tuned on standard LIBERO, not an in-domain LIBERO-X training result. OpenVLA/ACT's LIBERO-Plus and LIBERO-Pro runs are also zero-shot transfer (their checkpoints were fine-tuned/trained only on standard LIBERO); their standard-LIBERO runs are same-distribution. See the GitHub repository's docs/*-EVAL_REPORT.md and analysis-report/cross_policy_report_zh.md for the full per-model analysis, including OpenVLA's 0/2,000 catastrophic failure on the LIBERO-Pro swap(position) perturbation and ACT's language-blind architecture (documented so its Language Instructions/lan(semantic) perturbation scores aren't misread as language robustness).

Licensing

These are generated rollouts, not a redistribution of model weights or simulator assets. Users must follow the upstream licenses and citation terms for OpenPI, OpenVLA, lerobot/ACT, LIBERO, LIBERO-Plus, LIBERO-Pro, and LIBERO-X before using or redistributing the videos.