THU-SIGS-EILAB/coevo_cycle2_recap

CoEvo Cycle 2 RECAP teacher (self-rollout + HIL Cosmos), epoch 5

pi0.5 World Pilot policy after RECAP SFT on Cycle 2 self-rollouts with Cosmos-Policy takeover (automated HIL). Epoch-5 checkpoint (step 8975) of the run cycle2_step80_self_hil_recap_trainsplit_teacher, seed 42.

• Initialized from: Cycle 2 step-80 student (coeva_cycle2_acopd_80)
• Data: train split of the self-rollout + HIL collection, 401 LIBERO-10 episodes (114,883 frames). The other 99 episodes are held out.
• Labels: RECAP advantage, raw-progress top-30% positive, plus intervention frames forced positive (recap_n16_raw_top30_iv_exact_all)
• Conditioning: future-latent and value steering from Cosmos-Policy (30% joint condition dropout in training), RECAP CFG (guidance scale 1.5, unconditional prob 0.1)
• Training: 1795 steps per epoch, global batch 64, LR 2.5e-5 constant after warmup, full fine-tune (not LoRA). This checkpoint is the end of epoch 5.

Model weights only; the optimizer state is not included.

LIBERO-10 evaluation (500 episodes per seed)

Inference setting	Seed 42	Seed 43	Seed 44	Mean
Default: Cosmos z/v steering on, RECAP CFG 1.5	81.4%	81.8%	79.8%	81.0%
z/v off, RECAP positive prompt only (CFG 1.0)	82.8%	81.6%	84.2%	82.9%
z/v off, RECAP CFG 1.5	82.0%	79.2%	78.8%	80.0%
z/v off, plain task prompt (no RECAP)	81.8%	76.2%	77.4%	78.5%
Baseline: Cycle 2 step-80 student, plain	64.4%	65.8%	67.0%	65.7%

"z/v off" disables Cosmos future-latent and value steering, so Cosmos-Policy is not queried at inference. Per-run results, including the resolved eval configs, are in eval/.

Files

File	Contents
full_weights.pt	State dict, 877 tensors, 4.17B params (bf16 + fp32)
training_config.yaml	Resolved training config
physical-intelligence/libero/norm_stats.json	Action/state normalization stats
eval/*.json	Evaluation results

Code: branch RECAP of RLinf_opd. Training config: examples/sft/config/cycle2_step80_self_hil_recap_trainsplit_6ep.yaml.