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 0.3 s. Both arms predict 1.5 s of actions, plus the past actions in their history window. We ported the authors’ DINOv3 encoder, cross-attention U-Net and double encoder from their released code, and verified exact numerical parity with it. Both arms use this architecture and the same training recipe, so they differ only in history length and execution horizon. Adaptations to our platform:

• The method replans every 0.3 s instead of 0.2 s, because observation delay and inference together take about 220 ms on our system.
• Both arms run asynchronously on the station’s blend runtime, which crossfades into each new plan over 400 ms. At this latency the paper’s 180 ms interpolation never takes effect. Image embeddings are cached between replans.
• Both arms use real-time chunking at serving time, which steers each new plan toward the actions already committed. We adapted it from flow matching to the policy’s 10-step DDIM sampler; the paper does not use it.
• Actions run at 30 Hz instead of 10 Hz, with all horizons matched in seconds.
• Actions are per-step end-effector pose deltas anchored to the robot’s commanded pose, with absolute gripper widths. The commanded pose is also a model input. In training the anchor is the previous recorded action, because our datasets do not store the commanded pose. The paper does not specify its delta reference.
• The control uses the method’s architecture and is trained for 200k steps, instead of the paper’s simpler single-encoder baseline trained for 400k steps. Both arms train at batch 48 with the authors’ EMA.
• Checkpoints are selected by validation DDIM MSE among the 5 best, without hardware screening.
• Evaluation covers 4 tasks with 30 paired placements per arm.