THU-SIGS-EILAB/coevo_cycle1_offline_recap

CoEvo Cycle 1 Offline RECAP

PI0.5/OpenPI policy trained with exact-rank offline RECAP using self-rollouts and World Pilot conditioning.

Files

• actor/model_state_dict/full_weights.pt: inference checkpoint, training step 5595
• physical-intelligence/libero/norm_stats.json: required LIBERO normalization statistics
• eval_result.json: raw evaluation output
• method_and_results.md: experiment details

Result

LIBERO-10 evaluation, seed 195, 50 trials per task:

64.2% success (321/500), compared with 59.4% (297/500) for the CoEvo baseline (+4.8 percentage points).

Per-task success: 86, 68, 58, 94, 62, 48, 40, 76, 48, 62%.

This is a single evaluation seed; task 6 remains a failure case.

Download

hf download THU-SIGS-EILAB/coevo_cycle1_offline_recap \
  --local-dir checkpoints/coevo_cycle1_offline_recap

RLinf usage

Set MODEL_DIR to the downloaded directory and run:

python evaluations/eval_pi05_libero10_world_pilot.py \
  --config-name pi05_libero10_teacher_e3_plus \
  runner.ckpt_path="$MODEL_DIR/actor/model_state_dict/full_weights.pt" \
  rollout.model.model_path="$MODEL_DIR" \
  rollout.model.openpi_data.norm_stats_path="$MODEL_DIR/physical-intelligence/libero/norm_stats.json" \
  rollout.model.openpi.world_pilot.use_recap_cfg=true \
  runner.eval_seed=195

The evaluation additionally requires LIBERO, RLinf OpenPI dependencies, and local Cosmos Policy/Predict2-2B model paths; see method_and_results.md for the full configuration.