CTS-MoE: Implicit Terrain Adaptation via Mixture-of-Experts for Perceptive Locomotion

2026-06-17 · Source: Artificial Intelligence · Field: Technology & Digital — Robotics & Autonomous Systems, Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

CTS-MoE is a novel system designed for perceptive legged locomotion, enabling robots to adapt implicitly to discontinuous terrains like stairs, gaps, and obstacles. Addressing the tension in multi-task reinforcement learning between shared behaviors and conflicting rewards, CTS-MoE integrates a dense mixture-of-experts (MoE) actor with perception-based gating to compose shared behaviors. It also employs a multi-critic with task-specific value heads to prevent value interference. The model is trained end-to-end using a single-stage concurrent teacher-student setup, which manages partial observability and bypasses sequential distillation, with task labels exclusively used during training. During deployment, the system's routing mechanism relies solely on perception, eliminating the need for external high-level selectors or terrain classifiers. Experiments conducted on a Unitree Go1 robot, both in simulation and on hardware, demonstrated superior task-aware specialization, achieving lower tracking error and higher success rates compared to monolithic baselines across various seen and unseen terrains.

Key takeaway

For robotics engineers developing adaptive locomotion systems, CTS-MoE offers a robust approach to overcome challenges on discontinuous terrains. You should consider integrating a perception-gated Mixture-of-Experts architecture to achieve implicit terrain adaptation without explicit terrain classifiers. This method improves task-aware specialization and reduces tracking errors, enabling your robots, like the Unitree Go1, to navigate complex environments more reliably and with higher success rates than traditional monolithic policies.

Key insights

CTS-MoE uses perception-gated Mixture-of-Experts and multi-critics for adaptive legged locomotion on discontinuous terrains.

Principles

• Balance shared behaviors with task-specific needs.
• Perception-based gating enables implicit terrain adaptation.
• Multi-critic prevents value interference in multi-task RL.

Method

CTS-MoE trains a dense MoE actor with perception-based gating and a multi-critic with task-specific value heads end-to-end in a concurrent teacher-student setup, using task labels only during training.

In practice

• Deploy MoE for implicit terrain adaptation.
• Use multi-critics in multi-task RL.
• Test on Unitree Go1-like robots.

Topics

• Legged Locomotion
• Mixture-of-Experts
• Reinforcement Learning
• Terrain Adaptation
• Unitree Go1
• Multi-task Learning

Best for: Research Scientist, Robotics Engineer, AI Scientist, Machine Learning Engineer

Related on AIssential

• LoComposition: Terrain-Adaptive Energy-Efficient Quadruped Locomotion without Gait Priors — Decoupling quadruped locomotion control into distinct mechanisms enables efficient, terrain-adaptive behavior without explicit gait priors.
• Robotic Policy Adaptation via Weight-Space Meta-Learning — WIZARD enables robotic VLA policy adaptation via weight-space meta-learning, generating LoRA parameters from video and language without fine-tuning.
• Mind Your Steps: A General Learning Framework for Accurate Humanoid Foothold Tracking — The framework enables terrain-agnostic 3D humanoid foothold tracking by dynamically sampling goals and using a robust target representation for real-world transfer.
• Metis: A Generalizable and Efficient World-Action Model for Autonomous Driving and Urban Navigation — Decoupling video generation from action prediction in WAMs, using a Mixture-of-Transformers and asymmetric attention, improves efficiency and generalization for autonomous navigation.
• Learning Spatiotemporal Sensitivity in Video LLMs via Counterfactual Reinforcement Learning — Video LLMs need counterfactual training to develop genuine spatiotemporal sensitivity beyond static shortcuts.
• Tool-Aware Optimization with Entropy Guidance for Efficient Agentic Reinforcement Learning — TAO-RL stabilizes agentic reinforcement learning by filtering tool-use trajectories and guiding exploration with entropy for efficient LLM policy optimization.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.