Robotics · Vision · Learning
Ran Cheng
Foundation models for robots·VLA·embodied AI·robot learning
I'm leading foundation-model research on robot learning — and I write deep, interactive explainers about how these systems work.
About
I'm the Research Executive Director at Primebot, where I lead a team building robotics foundation models — VLA, world models, reinforcement learning powered by universal reward functions, and memory & continual learning. Previously, at Ant Group, I led pretraining on ten-thousand-GPU clusters, post-training, and real-robot reinforcement learning; before that, I led an R&D department at Midea and scaled robotics products to over one million production units; and at Huawei Noah's Ark Lab, I built scene reconstruction and world-model systems for autonomous driving. I hold an M.Sc. from McGill University's Center for Intelligent Machines (advised by Gregory Dudek and David Meger) and a B.S. from Tongji University.
Writing
all posts →- 2026-06-14 The Reachability of Steering: One Distribution-Level Law Behind Every VLA Trick
- 2026-06-14 Discarded Predictions: Rewriting Action-Chunking's Waste as Latent Distributions, Reality-Grounded Supervision, and Lipschitz Geometry
- 2026-06-10 Q-Guided Flow, From the Ground Up: Guiding a Flow Policy with a Value Function at Test Time
- 2026-04-02 Modern Hopfield Networks, Geometrically: From Wide Memory Basins to Attention
- 2026-02-08 Resonant Manifold Network - A Physics-Inspired Approach to Continual Learning
Research
publications →Selected papers in LiDAR perception, semantic scene completion, visual odometry, and RL for autonomous driving — see the publications page.