Humanoid Locomotion in 2025: What the Research Says

2025 was a landmark year for humanoid locomotion research. The field shifted from “can we walk?” to “can we walk anywhere, reliably, while doing something useful?” Here are the papers that pushed things forward.

The Sim-to-Real Gap: Narrower Than Ever

The biggest practical advance of 2025 was the dramatic reduction in the sim-to-real gap for locomotion policies. Three papers stand out:

Domain Randomization at Scale

Carnegie Mellon’s locomotion group published results showing that aggressive domain randomization — varying physics parameters like friction, link masses, and actuator response curves across 8,000+ simulation environments — produced policies that transferred to real hardware with minimal fine-tuning.

The key finding: the randomization distribution matters more than its magnitude. Uniform distributions worked poorly; distributions that matched the structure of real-world variation (lognormal for mass, beta for friction) worked significantly better.

Terrain Curriculum Learning

ETH Zurich’s paper on curriculum learning for rough terrain showed that progressively increasing terrain difficulty during training — starting with flat ground and adding obstacles, steps, and slopes — produced dramatically more robust policies than training on diverse terrain from the start.

Their humanoid could traverse terrain types it had never seen during training, suggesting genuine generalization rather than memorization.

Proprioception-Only Policies

Perhaps the most practically significant result: Stanford’s group demonstrated that locomotion policies relying only on proprioceptive sensing (joint positions, velocities, IMU) — with no camera or depth sensor — could match the terrain generalization of vision-based policies for typical indoor and outdoor walking scenarios.

This matters enormously for deployment: proprioceptive sensors are cheap, reliable, and high-bandwidth. Vision adds complexity without proportional benefit for basic locomotion.

Model Predictive Control Returns

After several years of RL dominating locomotion research, 2025 saw a renaissance of model-based methods — specifically MPC.

The key paper: “Fast Humanoid MPC via Differentiable Simulation” showed that GPU-accelerated differentiable physics could run MPC at 200Hz — fast enough for real-time control of a full-size humanoid. The approach combines the interpretability and constraint satisfaction of MPC with the speed previously only achievable with learned policies.

What’s Still Hard

Despite the progress, several problems remain largely unsolved:

Fall recovery — most systems fail catastrophically when pushed beyond their training distribution. Recovering from falls in a generalizable way is an open problem.

Energy efficiency — current humanoids use 5-10x more energy than humans for equivalent locomotion tasks. This is primarily a hardware problem, but control strategies matter too.

Upper-body coordination — maintaining balance while performing meaningful upper-body tasks (carrying objects, opening doors) remains significantly harder than locomotion alone.

Implications for Open Source Projects

The most actionable takeaway for the open source community: the algorithms are not the bottleneck. The locomotion RL stack is well understood and well implemented in libraries like legged_gym and humanoid_gym. The bottleneck is hardware — specifically, low-cost actuators with good torque density and the ability to survive repeated falls.

This is where open source can contribute: not by reinventing RL algorithms, but by building the hardware platforms that let more researchers run experiments cheaply.