2026-09-05
A Science Robotics survey traces humanoid control from models through sim RL to generative methods, calling the mix physics-guided generative intelligence. Full text not retrieved.
Humanoids are supposed to match human agility, robustness, and expressivity in an anthropomorphic body. Locomotion control has been the bottleneck. Over a decade the field moved from engineered stability through learning to generation. The literature is scattered, and it is hard to see how optimization, learning, and predictive reasoning should fit together next.
Science Robotics published the review Evolution of humanoid locomotion control on 19 August 2026. Authors include Yan Gu, Guanya Shi, Fan Shi, Aaron Ames, Hao Su, and Koushil Sreenath. The journal site is paywalled and behind Cloudflare; the PDF fetch returned a login page, and Unpaywall has no open copy. What follows is from the abstract only. Method detail, the citation map, and any numbers live in the unread full text.
This is a survey, not a new controller. The abstract organizes the field as three control generations plus three principles that cut across them.
The generations are chronological. Classical model-based methods: explicit dynamics, reduced templates, trajectory optimization, constraint satisfaction. Stable and analyzable, weak on messy terrain and contact switching. Reinforcement learning in large-scale simulation: parallel rollouts, policies that emit joint commands, generalization via randomization and curricula, at the cost of blurrier safety margins. Generative models: learn a motion distribution, then sample or guide whole-body behavior that adapts to a scene, moving from tracking demonstrations to synthesizing toward a goal.
The three principles are meant to pin those generations to one scorecard. Physics-based modeling: whatever the policy, execution still has to sit on real dynamics. Constrained decision-making: contact, friction, joint limits, and balance are a feasible set that data cannot wish away. Adaptation to uncertainty: model error, sensing noise, and pushes are the default at deployment.
The meeting point is called physics-guided generative intelligence: optimization supplies constraints, learning supplies coverage, generative models supply prediction and sampling. How the full paper splits case studies, and how it draws the line between generation and imitation, is not in the abstract.
There is no new comparison table. The abstract's claim is qualitative: humanoid control is at a turning point and is converging on that unified paradigm. Open problems are grouped as safety, accessibility, and human-level capability. Accessibility likely covers both open stacks and the cost of reproducing methods outside a few labs. Human-level capability is the step from walking under disturbance to collaborating with people as generalists in the open world.
The closing line is worth keeping: a shift from engineered stability to intelligent autonomy, laying groundwork for humanoid generalists that operate safely, collaborate naturally, and extend human capability outdoors. No success rates, no robot names, no single paper marked as the hinge. If those numbers exist, they are behind the paywall.
For people building humanoids, the value is positioning. If a lab is still training a slightly more stable walking policy, the survey's bet is that the next contest is generative whole-body behavior, and that generation has to be held by physics rather than video priors alone. Listing safety, accessibility, and human-level skill as open challenges is a reminder that a paper number is not a factory or a home. Because the body was not retrieved, this review cannot be used as a shopping list for controllers. The only firm claim is the narrative: optimization, RL, and generation are written as converging, not as replacements.
The binding limit is access. Depth here is abstract-level; the page will flag that full text was not retrieved. Surveys also select. Naming safety and accessibility in the abstract does not mean the body spends equal space on failures. Physics-guided generative intelligence is a programmatic label. The abstract does not say whether it requires model-predictive control or only a projection step after sampling. Calling generalist humanoids groundwork gives no timeline and no falsifiable metric. The review has to be read in full; the abstract is not a roadmap.