CodeWorld-Safe-LIBERO-Safety-pi05-GRPO-rollout

# CodeWorld-Safe SC-GRPO bundle (LIBERO-Safety, pi0.5, dual-view Wan 41k) Everything needed besides the three large model checkpoints (see "Also needed") to run safety-constrained GRPO of pi0.5 inside the frozen action-conditioned Wan world model. Code: github.com/Warrenustc1958/CodeWorld-Safe (`rlinf/envs/sim/world_model/safe/`, `examples/embodiment/config/libero_safety_scgrpo_openpi_pi05.yaml`, `examples/embodiment/run_scgrpo.sh`, `toolkits/codeworld_safe/scgrpo/`). ## keyframes_v1 (GRPO initial states) 47,658 real keyframes from 8,867 Wan-seen train episodes (SAM roles cost_train / h_train / dev; no diag or fresh data), chunk boundaries t0 (multiple of 8) with a success or safety event within 64 steps, plus 5 % random. - anchor.npy uint8 [N, 512, 256, 3] frame 0, agentview over wrist, policy (180-rotated) view - memory.npy uint8 [N, 4, 512, 256, 3] frames t0-3..t0 - meta.npz mem_actions [N,4,7] (env actions that produced the memory frames), states [N,8] (pi0.5 proprio eef_pos + axis-angle + gripper_qpos at t0), uid, group, tcell, suite, task_id, instruction, t0, kind (0 rest / 1 success / 2 event / 3 both), quadrant, role Large arrays are split: `cat anchor.npy.part_* > anchor.npy; cat memory.npy.part_* > memory.npy; sha256sum -c SHA256SUMS`. ## scoring - sam_cost_all_d32/best.pt SAM safety cost head on Wan latents (3 channels c_obj / c_robot / c_grasp) - channel_mask.json per task cell, which cost channels have grader rules - state_reg_v1/state_reg.pt CNN dual frame -> pi0.5 proprio (used for imagined steps) - rm_v3_agent_w41k/best.pt success reward model (agent view), success = 2 consecutive frames >= 0.2 ## eval/fresh_init_v1 20 fresh init states per task (40 tasks), unseen by the policy and Wan; real LIBERO-Safety evaluation with the grader (`toolkits/codeworld_safe/scgrpo/eval_real.sh`, loaded via `env.eval.init_states_root`). ## Also needed (not in this repo) - Wan dual-view DiT ema-step-41000: ModelScope Warrenustc1958/Wan-CodeWorld-Safe-pi05-dual-41000 - Wan2.2 VAE (Wan2.2_VAE.pth): HuggingFace RLinf/RLinf-Wan-LIBERO-Spatial - pi0.5 LIBERO-Safety RLinf checkpoint (model.safetensors + assets/lerobot/norm_stats.json) - Python env: RLinf openpi venv + diffsynth-studio on PYTHONPATH, plus modelscope==1.20.0, ftfy==6.3.1, seaborn==0.13.2; LIBERO-Safety (for real-env evaluation only)