libero40-trqc-checkpoints

LIBERO-40 task-relevant auxiliary-supervision checkpoints

This public repository archives the three checkpoints intentionally preserved at the end of our four-suite LIBERO study. The weights are modified/fine-tuned derivatives of the OpenPI pi0.5 base model.

Adopted result

The adopted checkpoint is checkpoints/pi05_libero40_trqc_supplemental115_from30000_seed42/3000.

• Four-suite main training: 30,000 optimizer updates
• Continuation: 3,000 updates on the original 1,693 LeRobot episodes plus 115 replay-validated supplemental episodes (1,808 episodes total)
• Auxiliary supervision: task-relevant Geometry + Semantic + Motion
• Coefficients: Geometry 0.05, Semantic 0.01, Motion 0.05
• Grounding supervision: disabled for this result
• Global batch: 256; training seed: 42; BF16; EMA disabled
• Formal evaluation: seed 7, 50 trials per task, 40 tasks, 2,000 rollouts
• Overall result: 1,957 / 2,000 = 97.85%

Per-task successes out of 50:

• Spatial T0-T9: 50, 50, 50, 50, 50, 50, 50, 50, 50, 50
• Object T0-T9: 50, 49, 49, 49, 48, 50, 48, 50, 50, 50
• Goal T0-T9: 50, 49, 50, 45, 50, 50, 50, 50, 50, 45
• LIBERO-10 T0-T9: 48, 50, 49, 50, 50, 46, 47, 49, 41, 45

Evaluation used 224x224 policy input, 10 flow steps, replanning every 5 environment steps, 10 initial wait steps, and 16 evaluation shards. Maximum episode lengths were 220 / 280 / 300 / 520 for Spatial / Object / Goal / LIBERO-10 respectively. Full task descriptions and counts are in results/supplemental115_step3000_summary.json.

Archived checkpoints

The two continuation checkpoints intentionally contain an EVALUATION_ONLY marker because their optimizer states were removed during local archival. The main step-30000 checkpoint is the retained exact-resume checkpoint.

Some provenance fields in metadata.pt contain absolute paths from the original machine. They document the run but are not portable configuration paths. Relocate datasets and auxiliary caches through the experiment release configuration when using the checkpoints on another machine.

Use SHA256SUMS to verify every checkpoint component after download.

Terms and attribution

These checkpoints contain modified Gemma/PaliGemma-derived parameters. Gemma is provided under and subject to the Gemma Terms of Use included in LICENSE_GEMMA.txt; downstream users must also comply with the incorporated Gemma Prohibited Use Policy. The required distribution notice is included in NOTICE.

The OpenPI software is distributed under Apache License 2.0; a copy is included as OPENPI_APACHE_LICENSE.txt. LIBERO benchmark data are not redistributed in this checkpoint repository.