Stanford's MobileVISTA fixes mobile manipulation's pose brittleness with generative data augmentation
jiajunwu_cs · x · 2026-10-08
Stanford University and Toyota Research Institute introduce MobileVISTA, a data generation framework for mobile manipulation. End-to-end policies trained on demos from a single base pose break when navigation leaves the base just a few centimeters off, pushing egocentric observations and end-effector trajectories out of distribution.
MobileVISTA jointly (1) augments egocentric visual observations with generative perturbations across poses and (2) retargets actions to compensate for base pose changes. Unlike prior work, it targets egocentric platforms like humanoids where the camera rides the actuated kinematic chain, and requires no extra demo collection or trained generative model.
Tested in simulation on humanoid and bimanual tasks and on a real Galaxea R1 Pro, policies trained on MobileVISTA-augmented data show markedly better robustness to previously out-of-distribution poses, with the largest gains on humanoids.
More from Embodied
- RLHND turns video diffusion models into physically grounded hand trackers for robot learning — RLWRLD · 2026-10-08
- NEEDLEwork stitches suboptimal robot demos into better training data — SongShuran · 2026-10-08
- Researchers compile a larval fish brain into a simulated circuit running 26x faster than spiking models — mtizard · 2026-10-08
- US Treasury issues first outbound investment fine: $200k over $92k Chinese robotics AI deal — pstAsiatech · 2026-10-08
- Palmer Luckey: I Believe in Humanoid Robots, But Won't Build Them Myself — PeterDiamandis · 2026-10-08
- Adding Motion Sense to VLAs With Frozen SAM 2 Lets Robots Catch Moving Bottles — nikhilaravi · 2026-10-08