Embodied AI for Humanoid/Legged Robotics

Our Approach

1. Foundation models for perception and control

   VLA models that connect natural-language instructions and semantic understanding with low-level motor control for locomotion and manipulation.

2. Bio-inspired embodied intelligence

   Reinforcement learning, model predictive control and sim-to-real transfer for agile, stable walking and climbing on diverse terrain.

3. Multimodal learning for humanoids

   RGB-D, IMU, tactile and force sensing combined into world models for whole-body planning and contact-rich manipulation.

Target applications

• Logistics and warehousingRobots that carry, sort and deliver goods on spoken instructions
• Assistive roboticsGuide-dog robots and service robots for elderly care
• Urban maintenanceLegged inspection in places wheels cannot reach
• Space explorationQuadrupeds and humanoids for lunar and planetary terrain
• Search and rescueLegged robots for disaster-stricken environments

Research in Action

Systems we have built and tested with partners. Open a card for the problem, our approach, results and photos.

Quadruped Guide-Dog Navigation Demonstration

A guide-dog robot for visually impaired users that navigates crowded campus paths and explains itself by voice.

• A → B campus demonstration
• Voice human–robot interface
• Joint lab PolyU–LinXAI

Problem, approach & results

ProblemA guide-dog robot must navigate crowded campus paths safely, explain itself by voice and stop safely when unsure.

Approach

• Perception, path planning, learned locomotion control, interaction and validation in one pipeline
• Real-time obstacle detection and avoidance; voice commands and spoken stop explanations
• Point-A-to-B demonstration on the PolyU campus

ResultsSystem design, implementation, demonstration and documentation.

Quadruped VLA for Factory Logistics

Quadrupeds that carry loads safely on spoken instructions, trained against worst-case disturbances.

• Adversarial robust training
• VLA language-guided missions

Problem, approach & results

ProblemLearned legged controllers fail under heavy payloads, joint faults and slippery floors in real factories.

Approach

• An “antagonist” policy that finds worst-case forces, joint faults and friction changes
• A hardened “protagonist” locomotion policy trained against it
• Vision-language-action model for natural-language transport missions

ResultsBuilds on the LinXAI quadruped platforms.

Quadruped Vision-Task-Action Model for Lunar Exploration

Legged robots that understand scientific tasks on the Moon and walk robustly on loose regolith.

• VTA vision-task-action model
• 2026 Deep-Space Robotics report

Problem, approach & results

ProblemWheeled rovers get stuck in craters; legged robots can go further but need task understanding and robust gaits.

Approach

• Vision-task-action model linking scientific goals to locomotion policies
• Robust gaits for loose regolith and slopes
• Lunar-surface simulation and terrestrial analogue tests

ResultsContribution to PolyU's Deep-Space Robotics Report (2026); TPC Chair of the APRIM 2026 satellite meeting.

Bio-inspired Spinal System for Humanoid Robots

A flexible, bio-inspired spine that helps humanoid robots balance on uneven, low-gravity terrain.

• Spine novel bio-inspired design
• Prototype humanoid robot
• Space exploration target application

Problem, approach & results

ProblemRigid one-piece torsos limit humanoid flexibility and balance on uneven, low-gravity terrain such as the Moon or Mars.

Approach

• Spine design that mimics human vertebral mechanics
• 3D-printed lightweight alloy and polymer vertebrae
• Data-driven MPC with a custom solver for the spine's nonlinear dynamics

ResultsSpine design, a functional prototype and robotics-journal papers.

Multi-robot Collaborative Operations in Lunar Areas for Regolith Processing

Navigation and formation control for teams of robots processing regolith on the Moon, without any GNSS.

• 0.30 m positioning RMSE
• Tianwen-3 team visit, Dec 2024

Problem, approach & results

ProblemLunar regolith tasks need precise, safe multi-robot operation on rugged terrain without any GNSS.

Approach

• Factor-graph positioning shared across the robot team
• Leader–follower formation control with MPC
• Prototype tests on rugged analogue terrain

ResultsVisit by the Tianwen-3 delegation in December 2024.

Video Demonstrations

Selected Publications

• Integrated Planning and Control on Manifolds: Factor Graph Representation and Toolkit.
  Yang, P., Wen, W., Yang, R., Zhang, Y., Hu, J., Chen, Y., Xiao, N., Zhao, J.
  IEEE International Conference on Robotics & Automation (ICRA), 2026.
• Unified Sufficient Conditions for Exact Convex Relaxation of Nonconvex Optimal Control Problems.
  Yang, R., Wen, W., Yang, P., Zhao, Z. and Huang, F.
  IEEE Transactions on Aerospace and Electronic Systems, 2025. (IF: 5.7, JCR Q1)
• EIRM-RL: Epistemic Integrity Risk Monitoring Inspired Safe Reinforcement Learning for Trustworthy Autonomous Navigation.
  Zhang, Y., Wang, Y., Wen, W.
  IEEE Internet of Things Journal, 13(2), 3500-3512, 2025. (IF: 8.9, JCR Q1)
• Learning Safe, Optimal, Real-Time Flight Interaction with Deep Confidence-enhanced Reachability Guarantee.
  Zhang, Y., Wang, Y., Yan, P., Wen, W.
  IEEE Transactions on Intelligent Transportation Systems, 2025. (IF: 8.4, JCR Q1, Citations: 2)
• Tightly Joined Positioning and Control Model for Unmanned Aerial Vehicles Based on Factor Graph Optimization.
  Yang, P., Wen, W.*, Bai, S., Hsu, L.T.
  IEEE Transactions on Vehicular Technology, 2025. (IF: 7.1, JCR Q1, Citations: 4)
• Online Dynamic Model Calibration for Reliable Control of Quadrotor Based on Factor Graph Optimization.
  Yang, P., Wen, W.*, Bai, S., Hu, J.
  IEEE Transactions on Aerospace and Electronic Systems, 2025. (IF: 5.7, JCR Q1, Citations: 2)
• Safe-Assured Learning-Based Deep SE(3) Motion Joint Planning and Control for UAV Interactions with Dynamic Environments.
  Zhang, Y., Wen, W., Yan, P.
  IEEE ITSC 2024. (Citations: 4)
• Tightly Joining Positioning and Control for Trustworthy Unmanned Aerial Vehicles Based on Factor Graph Optimization in Urban Transportation.
  Yang, P., Wen, W.
  IEEE ITSC 2023. (Citations: 7)

Full publication list →