e8: BONES-SEED full-set policy (100k iterations, 4×5090)

Headline policy of the reproduction: trained with the paper's observation layout on the full BONES-SEED set (129.9k clips) for 100k iterations on 4×RTX 5090 (7.1 h), see docs/EXPERIMENTS.md (e8) in the repository.

Appendix-F benchmark (whole clips, deterministic, no noise / DR): normalized return 0.930, 495 / 129,943 failures; 145 / 129,491 (0.11 %) after removing the acrobatics keywords the paper also filters (paper: >0.90, 25 / 129,514 on 1×H100 in 3 h).

Files:

• e8_boneseed_ckpt_0100000.pt (223 MB): world model, policy, critic, optimizer states, sampler statistics and the full training config. Loads with scripts/eval_benchmark.py --ckpt, scripts/play.py --ckpt, scripts/media/record_rollout.py --ckpt, or sm-train --resume. Policy observation: 93-d proprioception history + 680-d goal (K=10 future frames), no anchor observation.
• config.json: the run configuration (tyro TrainConfig).
• e8_boneseed_benchmark_full.json.gz: per-clip benchmark results for all 129,943 clips (name, frames, steps, normalized return, failed, fail step, per-term means); merge of the four GPU shards.

Code at commit ba61053. Requires the pinned environment in pyproject.toml / uv.lock (torch 2.14.1+cu130, mjlab 8bb2bc9).

---

复现的正式策略：论文观测口径、BONES-SEED 全量（12.99 万条）、四卡 RTX 5090 训练 10 万次迭代（7.1 h），详见仓库 docs/EXPERIMENTS.md 的 e8。Appendix F 全量评测回报 0.930、失败 145 / 129,491（0.11%）。检查点含世界模型、策略、critic、优化器状态与完整训练配置，可直接用于评测、回放、录制脚本或 --resume 续训；另附运行配置与 12.99 万条逐 clip 评测结果。