Zero-WAM fine-tuned on RoboICL
This private repository backs up completed training checkpoints from the RoboICL fine-tuning run initialized from robbyant-research/zero-wam-pretrain, revision 7040c4195df216c900334ef62d5fdcf05c0601aa.
The epoch 6 snapshot is at optimizer step 1,980. Its PyTorch checkpoint contains the LoRA and trained adapter parameters, optimizer and scheduler states, random states and training configuration. Inference requires the base model and the custom RoboICL adapter code. The checkpoint contains training state and should be loaded only from a trusted source.
Training uses four H200 GPUs, 16 directed demo/query pairs per GPU and global batch 64. LoRA rank is 106, with activation checkpointing and video/action/MCP losses. The demo supplies RGB task context. The query supplies observed RGB, measured state and 10D absolute world-frame hand/gripper action targets.
The source RoboICL release contains 3,000 recordings and 30,000 directed pairs. The local family holdout split retains 21,860 training pairs over 2,700 recordings and 260 validation pairs. Self-pairs and pairs crossing the family split are excluded. Equal-length batching presents 21,120 training pairs per epoch, giving 330 optimizer updates. The configured budget is 20 epochs and 6,600 updates.
No closed-loop success rate has been measured for epoch 6. A prior offline I/O test used the step 1 checkpoint.
Source data bundle SHA256:
75da8d99caebdc9bf7046456d3052e8ac17e0a26b79f7ae2e35cd919dfe2d8ee
Base code:
- https://github.com/robbyant-research/Zero-WAM
- https://github.com/imitator-game/The-Imitator-Game
- https://modelscope.cn/datasets/xieshanrui/RoboICL
The source model, code and dataset retain their respective licenses.