Locomotion

Locomotion covers the controllers, learning pipelines, and reference platforms behind legged, bipedal, and humanoid movement. It spans model-based whole-body control, RL-trained neural controllers, perceptive locomotion over rough terrain, and the GPU-parallel training stacks that make modern sim-to-real locomotion practical.

From an engineering standpoint, locomotion is the category where sim-to-real works — modern RL policies for quadrupeds and humanoids transfer reliably when training is done with massive parallelism and well-tuned domain randomisation. The interesting risks have moved up the stack: robustness to terrain and disturbances, perceptive integration with cameras and depth, and the safety envelope around fast, dynamic motion near humans.

When choosing between approaches, weigh platform (quadruped vs. biped vs. humanoid), simulator and trainer (Isaac Lab + RSL-RL is the current default), and perceptive vs. blind locomotion based on the terrain you actually need. Reference platforms and open codebases matter disproportionately here: most production locomotion stacks are forks of a small number of well-maintained repositories.

Start here

Learning to Walk in Minutes (ETH/RSL) is the canonical entry point: GPU-parallel RL with a working sim-to-real pipeline for quadrupeds. It pairs naturally with RSL-RL (the PPO trainer) and legged_gym (the environments), both listed below.

• Learning to Walk in Minutes (ETH/RSL) — Massively-parallel RL pipeline that trains quadruped locomotion policies from scratch in simulation.
• RMA — Rapid Motor Adaptation — Online adaptation method for robust real-world quadruped locomotion.
• ANYmal Parkour (RSL) — Perceptive locomotion enabling agile traversal of complex terrain.
• Cassie Bipedal Locomotion — Sim-to-real RL controllers for blind bipedal locomotion on Cassie.
• HumanPlus — Humanoid shadowing of human motion for whole-body control via teleoperation and RL.
• OmniH2O — Universal whole-body teleoperation and learning for humanoids.
• RSL-RL — Fast PPO implementation from ETH Zurich tuned for legged-robot RL on GPU simulators.
• legged_gym — Reference Isaac Gym/Isaac Lab environments for legged locomotion research.
• DeepMimic — Physics-based character control from motion imitation, foundational for agile RL locomotion.
• Walk These Ways — Learning framework for robust quadruped locomotion over varied terrain and command regimes.
• Rapid Locomotion via RL — Sim-to-real locomotion approach focused on high-speed deployment-ready policies.
• Agile and Dynamic Motor Skills — Learning-based control for dynamic maneuvers in legged robots under disturbances.
• Learning Quadrupedal Locomotion over Challenging Terrain — Classic ANYmal result demonstrating robust transfer from simulation.
• Isaac Gym — GPU-parallel simulation stack that accelerated modern locomotion training pipelines.
• MuJoCo Menagerie — Curated high-quality robot models for repeatable locomotion research.
• Real-World Humanoid Locomotion with RL — Sim-to-real RL controller for blind humanoid walking on a real bipedal platform (Science Robotics 2024).
• Periodic Reward Composition for Bipedal Gaits — Sim-to-real RL recipe producing walking, hopping, and skipping from a single Cassie policy (ICRA 2021).
• Robust Parameterized Bipedal Locomotion (Cassie) — RL controller covering a parameterised velocity/stance family on Cassie (ICRA 2021).
• Bipedal Soccer (DeepMind OP3) — Whole-body RL pipeline producing agile soccer skills on a small humanoid (Science Robotics 2024).
• Humanoid Parkour Learning — End-to-end perceptive RL enabling humanoid parkour over discontinuous terrain (CoRL 2024).
• Expressive Whole-Body Control — Whole-body controller producing expressive, human-like motion on a real humanoid (RSS 2024).
• ASAP — Sim-to-Real for Humanoid Whole-Body Skills — Delta-action correction closing the sim-to-real gap on agile humanoid skills (RSS 2025).
• HOVER — Versatile Humanoid Whole-Body Controller — Single neural controller covering multiple humanoid command modes (ICRA 2025).
• FLD — Fourier Latent Dynamics (MIT Biomimetics) — Periodic motion latent representation for learning agile, natural-looking legged locomotion (ICLR 2024).
• WASABI (Max Planck / Martius Lab) — Versatile skill learning for quadrupeds via unsupervised motion-prior discovery from unlabeled reference data.
• CASSI (Max Planck / Martius Lab) — Self-supervised adversarial imitation of unlabeled mixed motions for versatile quadruped skill control (ICRA 2023).