FastWAM T-Rex Egg v2 — RoboRAG-P (Prefilling ICL)
Final step-5,000 FastWAM checkpoint for the T-Rex Egg v2 task from
wushr-lance/VLA2Vec.
Task instruction: Pick up the white egg from the orange egg cup with the right hand and place it in the purple cup.
32x1024 reference Context Tokens are prefilled; the policy and host adapter are fine-tuned end to end from the final T-Rex Egg v2 baseline. This is a consolidated full-model inference checkpoint. The fixed EgoDex reference video, assignment receipt, and precomputed 32x1024 Context Tokens are included under reference/. Runtime inference loads reference/context_tokens.safetensors directly, so neither the embedding model nor video decoding is required.
Final validation
| Metric | Value |
|---|---|
| Optimization step | 5,000 |
| Release status | validated |
| Legacy convergence gate | passed |
TRAINING_DONE marker |
true |
| Validation loss | 0.312292 |
| Action L1 | 0.016011 |
| Action L2 | 0.001135 |
The repository contains only the final inference checkpoint, not intermediate checkpoints or optimizer state. It also includes the resolved training config, complete metrics, convergence receipt, exact normalization statistics, and task text embedding. This is an offline validation receipt, not a physical-robot success rate.
Reproduction code: yunzeliu/fastwam_icl
at commit ce76230cc4658ac8df6d3efe09c9f85e3f37a22a. The native .pt files use FastWAM checkpoint
format and retain the T-Rex 62-D action contract (32x62 action chunks).
Repository: YunzeLiu/fastwam-trex-egg-v2-roborag-p-5k.
中文说明
这是 T-Rex Egg v2 任务的 Prefilling ICL 最终 5,000-step checkpoint。仓库只上传
训练完成后的最终推理权重，不包含中间 checkpoint 和优化器状态；同时保存训练配置、
完整指标、收敛检查、归一化统计和任务文本特征。ICL 版本额外包含固定的人类参考视频
及预抽取的 32x1024 Context Tokens。 真机部署前必须保留 T-Rex 的 IK、碰撞、限速和
急停保护，并先进行断开执行器的输入输出检查。
Egg v2 data and initialization
Pinned dataset revision: d9057f25298097b5745bfd86af2727b544a4022e.
Subset: teleop_data/transfer_egg_v2/success; 100 successful episodes,
50,366 frames. Old Egg episodes are not mixed into training.
The robot transfers the egg from the orange cup to the purple cup.
Text context and normalization statistics were recomputed for v2.
Both stages use seed 42, global batch 128, and 16 GPUs. Baseline trains for 5,000 updates from Wan2.2/interpolated ActionDiT initialization at peak LR 1e-4. RoboRAG-P starts from this v2 baseline and trains for 5,000 further updates at peak LR 5e-5.
The fixed human reference matches right-hand egg grasp, transport and placement; its source egg carton and destination red plate differ from the robot cups. The baseline does not use this reference.
Weights are at checkpoints/step_005000.pt. Matching normalization statistics and
the new task text context are in artifacts/. This release has not been measured
for physical-robot task success.