ManiFlow

We ported ManiFlow’s DiT-X policy from the official code and compare it with the platform’s Diffusion Policy. ManiFlow conditions on 64 feature-map tokens from each of three cameras plus a state token, and samples actions in 2 steps. Both arms share the ImageNet ResNet-18 backbone, data, image augmentation, a 64-step horizon, the training budget and EMA weights. Differences from the paper:

• Ran the paper’s RGB comparison against Diffusion Policy, which it reports only in simulation, on real robots instead of its point-cloud setup against 3D Diffusion Policy.
• Used an ImageNet-pretrained ResNet-18 in both arms instead of training one from scratch.
• Conditioned on one observation frame instead of two, because the station sends one observation per request.
• Executed 32 steps of each 64-step prediction and blended successive plans on the station instead of temporal ensembling, assuming 30 Hz because the paper doesn’t state its control rate.
• Controlled absolute end-effector poses with a parallel gripper instead of joints and dexterous hands.
• Kept the platform’s Diffusion Policy preset for the control, with a wider U-Net and 16 DDIM steps, rather than the original real-robot configuration.
• Trained one seed per arm and evaluated its final checkpoint instead of averaging three seeds.