Yi Wei危伊

Diffusion and flow-matching visuomotor policies capture multimodal action distributions, but their multi-step sampling is too slow for fast control, and robot actions often live on a curved manifold that the policy has to respect. We present Riemannian MeanFlow Policy (RMFP), which learns the conditioned flow map of the probability path on the robot action manifold, trained with a flow-map consistency objective anchored by Riemannian Conditional Flow Matching. RMFP generates on-manifold action sequences in as few as one network evaluation, is competitive with prior work at lower sampling cost on spherical LASA, Push-T, Robomimic (Tool Hang, Transport), and Franka Kitchen, and runs on a real robot under imperfect sensing.

My contribution:

@article{bukhari2026rmfp,
  title   = {Faster Visuomotor Policy Learning on Action Manifolds via
             Riemannian MeanFlow},
  author  = {Bukhari, S. Talha and Garrett, Austin and Wei, Yi and
             Ni, Ruiqi and Kingston, Zachary and Bera, Aniket},
  journal = {arXiv preprint arXiv:2609.30127},
  year    = {2026}
}

Grasp synthesis is multimodal by nature. A good grasp is a distribution, not a point estimate, which makes generative models a natural fit, but their iterative sampling is too slow for time-critical robot control. We formulate MeanFlow on the product Lie group SO(3) × ℝ³, coupling an algebraic semigroup consistency condition with Riemannian Conditional Flow Matching so that the average velocity stays anchored to the data distribution. The model samples reliable grasps in ≤ 5 network evaluations, matching state-of-the-art diffusion and flow-based grasp generators on ACRONYM at millisecond-scale latency (up to 39× speed-up), and transfers directly to real-world grasping with no additional training or domain adaptation.

My contribution: I owned the real-robot side, turning the model into a working stack on a physical Franka R3. I built the calibrated eye-in-hand RGB-D pipeline with an Orbbec Femto wrist camera, diagnosed and fixed a ~30 mm hand-eye error caused by a frame-convention mismatch, designed a shared-correction multi-view ICP that cut validation alignment error from 11.6 mm to 10.2 mm, wrote the adapter mapping generated SE(3) grasps into the robot frame, and integrated collision-aware VAMP planning and validated end-to-end execution to roughly 1 mm / 0.13°.

@article{bukhari2026meanflowgrasp,
  title   = {Fast Generative Grasping via Lie Group-Constrained MeanFlow},
  author  = {Bukhari, S. Talha and Wei, Yi and Ni, Ruiqi and
             Kingston, Zachary and Bera, Aniket},
  journal = {arXiv preprint arXiv:2608.26076},
  year    = {2026}
}