ManiFlow

We ported ManiFlow’s DiT-X policy and its consistency flow training from the official code and compare it with the platform’s Diffusion Policy. ManiFlow samples actions in 10 steps and conditions on a single token that concatenates the robot state with every camera’s features. Both arms share the ImageNet ResNet-18 backbone, data, the paper’s image augmentation, a 32-step horizon and the training budget. 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.
• Fed DiT-X’s cross-attention one conditioning token instead of multiple visual tokens.
• Sampled in 10 steps instead of the 1 to 2 steps behind the paper’s headline result.
• 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.
• Predicted and executed 32-step action chunks, blended on the station, instead of 64-step chunks with temporal ensembling.
• Served the online weights in both arms instead of the EMA weights the official code evaluates.
• 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.