gulucaptain/zero-wam-roboicl-version-a

Robot-ICL + A

Version A adds three zero-initialized, gated cross-attention layers that allow Zero-WAM's action branch to read tokens from a robot demonstration. A frozen DINOv2 ViT-B/14 encodes eight RGB samples from the external camera. Each sample contributes a 2×2 grid of pooled features with time and space indices. Video ICL and the original MCP training objective are retained.

Training warm-starts from the Robot-ICL baseline at step 3,698, which uses Zero-WAM pretraining weights (Apache-2.0, revision 7040c4195df216c900334ef62d5fdcf05c0601aa). The RoboICL fine-tuning and demo input use robot data. The inherited Zero-WAM weights were pretrained with human demonstrations.

Files

• epoch-NNN/adapter.safetensors contains the trainable tensors, including LoRA, state/action heads, MCP adapters and A's attention layers.
• epoch-NNN/adapter.json records the optimizer step, parameter count and file checksum.

To run inference, load the base Zero-WAM model and attach these tensors with the matching Robot-ICL + A integration.

Training is configured for 20 epochs with 330 optimizer steps per epoch, four GPUs, and a global batch of 64. Epoch snapshots are published after the corresponding training checkpoint is saved. Closed-loop performance has not been measured for these snapshots.