yky666/g1d_experiments

G1D experiments — public model weights

代码、模型实现、训练/评测和部署说明统一在 GitHub g1d_experiments。本公开HF模型仓库统一存放权重与运行所需模型资产；无需私有访问权限。

模型	目录	状态
π0.5 LoRA subtask baseline7500	pi05_lora_baseline/7500	已发布、源文件校验通过
AFCE/C01 subtask 全量FT72500	afce_fullft/72500及afce_fullft/cache	已发布、源文件校验通过

AFCE完成100k训练后按200观测heldout composite loss选择72500（0.1550050788745284）；100000为0.15692223992198706。heldout用于选点，不是独立测试或真机成功率。π0.5 baseline为已有实机经验的7500权重，训练预算不同。

下载

hf download yky666/g1d_experiments --include "pi05_lora_baseline/**" --local-dir artifacts/models
# AFCE上传完成后下载：
hf download yky666/g1d_experiments --include "afce_fullft/**" --local-dir artifacts/models

可在GitHub ARTIFACTS.json中取不可变revision，然后为hf download添加--revision <commit>；或使用仓库的python scripts/download_artifacts.py --baseline --afce-fullft，自动固定revision并保持评测脚本目录结构。

输入输出与服务

• 三路当前RGB（主视角/左右腕）、原生16D实测state（双臂各7关节＋两个夹爪，输入补到32D）与subtask prompt。
• Baseline预测32×16动作；AFCE先预测E空间30×256，再由联合训练decoder输出30×16动作；30Hz是客户端执行频率，与chunk长度是不同概念。
• AFCE推理不需要DINOv3和E编码器；decoder初始化/normalization/manifest与norm_stats已附带。训练后decoder已在checkpoint参数内。
• Orbax/JAX完整推理params，不是Transformers格式或仅LoRA adapter；不含optimizer。
• 仍使用subtask指令/外部阶段切换；这是原子能力对比权重，不是三整任务40k的新实验。
• 同observation比较取双方前30步/native16；闭环需固定初始布局、控制频率和控制参数。

git clone --recurse-submodules https://github.com/HERMES-Unifyed-Robotic-Control/g1d_experiments.git
cd g1d_experiments
# 配好README规定的JAX/Flax/Orbax和tokenizer环境后：
python scripts/serve_c01.py --checkpoint /path/to/artifacts/models/afce_fullft/72500 --cache /path/to/artifacts/models/afce_fullft/cache --variant fullft --port 9023 --denoise-steps 10 --warmup

数据仍在独立的 humanoid_long_horizon_tasks dataset，代码以GitHub为准。

2026-10-18 checkpoint updates

Task / model	Path	Checkpoint status
food_sorting AFCE subtask LoRA	food_sorting/afce_lora/97500	Selected from100k training; inference params/assets
food_sorting chrysanthemum whole-task AFCE fullFT	food_sorting/afce_whole/chrysanthemum/2500	Selected additional2.5k from40k adaptation initialized at72.5k
pick_bucket pi0.5 LoRA	pick_bucket/pi05_lora/70000	Selected checkpoint from100k training; train_state included
pick_bucket pi0.5 LoRA	pick_bucket/pi05_lora/100000	Final complete checkpoint, not selected best; cursor and run metadata under provenance/
pick_bucket pi0.5 fullFT	pick_bucket/pi05_fullft/12500	Intermediate complete checkpoint; training continues toward100k

Original baseline7500, AFCE subtask fullFT72500 and its E cache remain at pi05_lora_baseline/7500, afce_fullft/72500, afce_fullft/cache. Best checkpoints are selected on the existing heldout split, not an independent test. Open-loop losses do not establish real-robot success.

Native state/action:16D (left arm7, right arm7, left gripper, right gripper); policy state padded to32D. AFCE horizon30; ordinary pi0.5 horizon32. Preserve each checkpoint's normalization, prompts and matched E assets. Files containing train_state alone do not establish resumability without the matching data contract/cursor/runtime.

hf download yky666/g1d_experiments --include 'pick_bucket/pi05_lora/70000/**' --local-dir ./g1d_models

See MODEL_INDEX.json for versioned paths and verification details.