Amazon FAR's PRISM Turns 4 Real Videos Into 256 Counterfactual Variants to Train Humanoid Robots

chris_j_paxton · x · 2026-10-01

Amazon FAR, with UC Berkeley, Stanford and CMU, unveiled PRISM (CoRL 2026), a real-to-sim-to-real pipeline for humanoid loco-manipulation. It expands 4 real demonstration videos into 256 counterfactual variants via video-to-video generation, recovers robot-object trajectories, and trains a single policy that generalizes across diverse objects, scales and layouts. The robot runs on onboard depth only—no mocap or reference motion—and handles novel real objects zero-shot, sidestepping costly demonstration collection. Paper, code and data are public.

Original post →

More from Embodied

Embodied channel →