NTU Proposes World Model Racing Agent to Explore Embodied Cognitive Limits
机器之心 · wechat · 2026-08-20
NTU's AutoManLab proposes a world-model-centric embodied agent architecture to address perception, decision-making, and control for autonomous racing at high speeds and in adversarial environments. The system uses a world model to simultaneously predict opponent interactions, vehicle dynamics, and feasible motion envelopes, optimizing strategies through a loop of simulation and real-world data.
Core Architecture:
- World Model Centricity: Integrates prediction of opponent behavior, self-response, and physical feasibility envelopes, replacing traditional serial perception-planning pipelines.
- Decision Rollout: Performs forward rollout of candidate trajectories via the world model before execution, scoring based on safety, feasibility, and speed.
- Physical Limit Constraints: Dynamically adjusts usable physical boundaries (e.g., tire grip, actuator limits) to prevent instability near limits.
Results:
- Achieved 260km/h real-world autonomous driving in A2RL, securing 4th place globally in 2025.
- Simulation: 88.3% interaction success rate, collision rate reduced to 5.8%, near 90% zero-shot transfer success on unseen tracks.
More from Embodied
- Matic Robot Adds Voice and Gesture Control: 'Clean This' Works in 75 Languages — lukas_m_ziegler · 2026-08-20
- SoftBank Leads $200M Round for Gravis to Automate Excavators — lukas_m_ziegler · 2026-08-20
- Ex-NVIDIA Team Raises $90M+ to Build World Models for Physical AI — lukas_m_ziegler · 2026-08-20
- Weekly Robotics Roundup: Updates from Unitree, Generalist, Gravis, and More — lukas_m_ziegler · 2026-08-20
- China Merchants Lab Unveils LiOS Infrastructure for Robotic Cloth Folding — 量子位 · 2026-08-20
- DaxAI's Rideable Robot Horse Qigi Now Available: X1 at $43K, Wheel-legged XS at $53K — Scobleizer · 2026-08-20