Research / Walking and balance

Can humanoid robots walk reliably yet?

Updated November 13, 2026 · 3 papers · How we rate evidence

Where things stand

On set lab courses, increasingly yes. Learned walking controllers trained in simulation now get real humanoids up stairs, across 70 cm gaps and over stepping stones; one November 2026 paper reports 98 successes in 100 hardware trials. What the papers do not show yet is long, unsupervised walking on a working floor, among people and clutter, for a full shift.

Most of the progress comes from training in simulation and transferring to the real robot without real-world practice. The newest systems add vision, so the robot can see stairs and gaps before it reaches them, and some keep a short memory of terrain it can no longer see.

Balance is being separated from motion. Instead of one controller doing everything, a physics model makes the robot follow a movement and a learned layer adds small corrections to keep it upright when it carries a load or is pushed.

Every result here is reported by the team that built it, on its own test course, over short trials. Treat success rates as a ceiling under good conditions, not a measure of everyday reliability.

The papers

Moderate evidencePhysical robot and simulationNot named in the abstract

DKWBC: a separate learned layer just for balance

DKWBC: Decoupling Motion Tracking and Balancing for Humanoid Whole-Body Control · Dingkun Wei et al. · arXiv 2611.09605 · Nov 6, 2026 · Preprint

What they did: Splits whole-body control in two: a physics model makes the robot follow a motion, and a learned policy adds small torque corrections so it stays balanced when standing, carrying a load or being touched.
Result: The authors report that one learned balance policy worked across different motion datasets, in simulation and on real hardware.
Caveat: The abstract reports no numbers and does not name the robot or the comparison methods.

Checked against: Abstract. Results as reported by the authors; not peer-reviewed unless noted.

Strong evidencePhysical robot and simulationNoetix E1 (23 joints)

STEP: seeing the terrain ahead, and remembering it

STEP: Spatiotemporal Training with Explicit Prediction for Perceptive Humanoid Locomotion · Youhao Lyu et al. (University of Science and Technology of China, Noetix Robotics and others) · arXiv 2611.08097 · Nov 5, 2026 · Preprint

What they did: A walking controller that uses a depth camera to predict the terrain ahead, and keeps track of terrain it can no longer see, such as the stairs behind it when it walks backward.
Result: 98 of 100 hardware trials succeeded across stairs, a 40 cm platform, a 70 cm gap, stepping stones and walking backward down stairs. In simulation it beat other ways of combining vision with body sensing.
Caveat: Short trials on a fixed course, run by the authors. Detailed simulation figures are in the full paper.

Checked against: Abstract and paper text. Results as reported by the authors; not peer-reviewed unless noted.

Preliminary evidencePhysical robot and simulationergoCub

Planning steps over uneven ground on ergoCub

Contact-Implicit Whole-Body Trajectory Optimization for Dynamic Humanoid Locomotion on Parameterizable Uneven Terrain · Roberto Mauceri et al. · arXiv 2611.05585 · Nov 3, 2026 · Preprint

What they did: Plans whole-body walking motions over uneven ground without scripting in advance where each foot will land.
Result: The authors describe preliminary feasibility tests in simulation and on the real ergoCub robot over non-flat terrain.
Caveat: No success rates or baselines are reported; the authors call the tests preliminary.

Checked against: Abstract. Results as reported by the authors; not peer-reviewed unless noted.

More research

All research Hands and handling objects Navigation Working with people Robot Rankings

How we pick and rate papers

We cover primary research that tests AI, perception, planning or control on a humanoid robot, or reports a distinct humanoid experiment. Product launches, funding news and promotional videos without a paper are left out.
Every paper links to its primary source. Most are arXiv preprints, which are not peer-reviewed; we say so, and add the journal or conference version when one is confirmed.
We check the numbers against the paper itself, and say whether we read the abstract or the paper text.
We separate physical-robot results from simulation, and note when a robot was teleoperated rather than acting on its own.
Strong means quantified tests on a physical humanoid with defined tasks and real comparisons. Moderate means limited hardware results or partial evidence. Preliminary means simulation-only or exploratory work.

Spotted an error, or a paper we should cover? Tell us, with a link to the paper.