Downloads · 30 days
0
YunzeLiu/fastwam-trex-cup-side-branch-icl-5k
fastwam-trex-cup-side-branch-icl-5k is a robotics model from YunzeLiu. Use it for the robotics task on the model card, and read the license before you ship it in a product. It is set up for fastwam.
Final step-5,000 FastWAM checkpoint for the T-Rex Cup task from wushr-lance/VLA2Vec. Task instruction: Separate the nested stack of red plastic cups using both hands, placing the cups apart on the table.
Downloads · 30 days
0
Access
Public
Updated Sep 1, 2026
Repo size
12.4 GB
Likes
0
Public
Click a slice to open those files.
.pt12.4 GB · 100%
From the Hugging Face model README
Final step-5,000 FastWAM checkpoint for the T-Rex Cup task from
wushr-lance/VLA2Vec.
Task instruction: Separate the nested stack of red plastic cups using both hands, placing the cups apart on the table.
The final Cup baseline is frozen; dense cross-attention side branches are injected at all 30 layers and only the new branch parameters are optimized. This checkpoint is a side-branch delta. Load the bundled base/step_005000.pt first. 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.
[!WARNING] This is an experimental completed step-5,000 checkpoint. The frozen-backbone and branch-gradient checks passed, but the legacy early/late training-loss gate failed: aggregate action loss did not improve enough: ratio=1.1909 > 1.0000; loss_action_left_arm did not improve enough: ratio=1.2477 > 1.0500; loss_action_left_hand did not improve enough: ratio=1.4553 > 1.0500; loss_action_right_arm did not improve enough: ratio=1.0971 > 1.0500; loss_action_right_hand did not improve enough: ratio=1.0626 > 1.0500. The original receipt is included unchanged; this release does not claim that the side branch improves task success.
| Metric | Value |
|---|---|
| Optimization step | 5,000 |
| Release status | experimental_gate_failed |
| Legacy convergence gate | failed (preserved verbatim) |
TRAINING_DONE marker | false |
| Validation loss | 0.306455 |
| Action L1 | 0.024621 |
| Action L2 | 0.002959 |
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 345f1bcf921b63970b7d1bc4705b7752af508e97. The native .pt files use FastWAM checkpoint
format and retain the T-Rex 62-D action contract (32x62 action chunks).
Repository: YunzeLiu/fastwam-trex-cup-side-branch-icl-5k.
这是 T-Rex Cup 任务的 Side-Branch 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。Side-branch 已完整训练到 5,000 step,但旧版早/末窗口训练损失门槛未通过;仓库原样保留失败 receipt,并标记为 experimental,不宣称任务成功率提升。 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。