FastWAM T-Rex Tablet — Prefilling ICL
Final step-5,000 FastWAM checkpoint for the T-Rex Tablet task from
wushr-lance/VLA2Vec.
Task instruction: Use your left hand to open one of the compartments in the small box, use your right hand to grasp the small tablet on the table, place the tablet into the box, and then close the box.
32x1024 reference Context Tokens are prefilled; the policy and host adapter are fine-tuned end to end from the final T-Rex Tablet 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 | 0.293798 |
| Action L1 | 0.016892 |
| Action L2 | 0.000966 |
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-tablet-roborag-p-5k.
中文说明
这是 T-Rex Tablet 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。