Open-Loop Execution

We reproduced the paper’s real-robot comparison: a Diffusion Policy with 2 history frames that executes 0.8 s of each plan, against one with 8 history frames that replans every 0.2 s. Both arms predict 1.5 s of actions, plus the past actions in their history window, and use the same network, so they differ only in history length and execution horizon. Adaptations to our platform:

• Both arms use the platform’s Diffusion Policy with a ResNet-18 encoder and a FiLM-conditioned U-Net at the paper’s widths, instead of the paper’s DINOv3 encoders with a cross-attention U-Net and double encoder.
• Actions are absolute end-effector poses instead of the paper’s delta actions.
• Actions run at 30 Hz instead of 10 Hz, with all horizons matched in seconds.
• Both arms run on the station’s synchronous runtime, and the server smooths each plan switch over about 150 ms, mirroring the paper’s 180 ms interpolation.
• Each arm serves its final checkpoint, because the cloud trainer kept only that one, instead of selecting by validation error as the paper does.
• Training uses a 95% random image crop, a batch of 64 and about 2% of episodes held out.
• Evaluation covers 5 tasks with 30 paired placements per arm.