coevo_cycle1_offline_recap_cosmos_rollout

CoEvo Cycle 1 Offline RECAP — Cosmos Rollout

This repository contains the consolidated PI0.5 model checkpoint trained with offline RECAP on Cosmos-Policy rollouts for LIBERO-10. It also includes the training/evaluation configuration, matched cohort manifest, and raw four-seed evaluation results.

Checkpoint

• File: full_weights.pt
• SHA-256: 636c0c55b76bcf28bd32eed4dd23a7abe4ecde635aee9b9a30c0348af2b2b221
• Training step: 5,340
• Initialization: checkpoints/coeva
• Format: consolidated PyTorch model state dictionary
• Action chunk and value lookahead: 16
• Training seed: 42
• Code commit: a2b11673 (feat(recap): compare matched rollout sources)

Distributed checkpoint shards, optimizer state, and other DCP training state are intentionally excluded. This is not a standalone Transformers model directory. The required LIBERO normalization statistics and a portable RLinf model fragment are included.

After downloading the repository, set PI05_RECAP_MODEL_DIR to its local path and compose rlinf_model_config.yaml with the appropriate RLinf LIBERO evaluation config. Online Cosmos latent/value conditioning is required to reproduce the reported RECAP results.

Experiment

RECAP was trained on 353 selected Cosmos-Policy episodes: 283 successes and 70 failures. The cohort was matched to the self-rollout source independently for every LIBERO-10 task and outcome. It contains 96,338 frames.

• Future condition: cached Cosmos latent z and scalar value
• Advantage target: 16-step Evo task-relative target
• Failure coefficient: 1.0
• Labeling: all successful frames are positive; the exact top 40% of failed frames per task are positive
• Threshold population: all selected episodes of the same task
• Training: 5,340 updates, global batch 64, micro-batch 16 per GPU, four GPUs
• RECAP CFG: positive suffix Advantage: positive, guidance scale 1.5
• Steering: latent and value enabled; action steering disabled

The exact files used for the experiment are under experiment/.

Evaluation

Each seed uses fixed ordered reset states, 50 trials per task (500 total), 32 parallel environments, and online Cosmos conditioning.

Mean success rate by task

Cosmos-rollout RECAP improves the four-seed mean by 13.55 percentage points over the teacher baseline. The largest gains occur on tasks 2, 5, and 9. Tasks 6 and 8 regress by 3.5 and 3.0 points, respectively.

The included existing self-rollout checkpoint result is contextual rather than a strictly controlled source comparison: it used 371 episodes and 5,595 updates. A 353-episode, 5,340-update matched self-rollout arm is prepared but has not yet been trained and evaluated.

Raw per-seed/per-task measurements and the detailed report are under results/.

Files

• full_weights.pt: consolidated model state dictionary only
• physical-intelligence/libero/norm_stats.json: required LIBERO action statistics
• rlinf_model_config.yaml: portable checkpoint and model configuration
• experiment/training_base.yaml: shared matched training configuration
• experiment/training_cosmos_rollout.yaml: Cosmos source configuration
• experiment/evaluation_cosmos_rollout.yaml: four-seed evaluation config
• experiment/cohort_manifest.json: exact selected episodes and counts
• results/cosmos_multiseed_results.json: Cosmos per-seed/per-task results
• results/cosmos_multiseed_summary.json: Cosmos aggregate results
• results/baseline_self_multiseed_results.json: comparison measurements
• results/experiment_report.md: protocol, caveats, and analysis