Science Robotics survey: humanoid control converges toward physically grounded generative intelligence

量子位 · wechat · 2026-09-05

A Science Robotics review by researchers from Purdue, CMU, NUS, UC Berkeley, NYU, Caltech and Meta charts decades of humanoid locomotion control and argues the field is converging on "physically grounded generative intelligence."

Three stages

Classical methods: reduced-order models (LIP, ZMP), trajectory optimization and MPC that write dynamics and constraints explicitly into the problem;
Learning-based: RL built on large-scale physics simulation (Isaac, MuJoCo) — really an relocation of optimization from online solving into offline-trained policies; domain randomization maps to robust control, hierarchical learning to cascaded control;
Emerging: diffusion policies, VLA models and world models entering the loop, shifting along five axes: discriminative→generative, unimodal→multimodal, single-task→multi-task, locomotion→locomotion-manipulation, offline training→test-time adaptation.

Key arguments

The central question for VLA is which layer it should control: most systems (Helix, NVIDIA GR00T) stay hierarchical; unified whole-body control is more ambitious but harder to validate;
Generated trajectories must satisfy contact, friction and torque constraints to count as control;
Evaluation should track falls and rare failures, not average success; data — especially rare-failure and contact-rich interaction data — remains the bottleneck.

Co-first authors: Yan Gu (Purdue), Guanya Shi (CMU), Fan Shi (NUS); senior authors Hao Su and Koushil Sreenath.

Original post →

Science Robotics survey: humanoid control converges toward physically grounded generative intelligence

More from Embodied