RAY-TOLD: Ray-Based Latent Dynamics for Dense Dynamic Obstacle Avoidance with TDMPC
Blends short-term physics with long-term RL to avoid collisions.
RAY-TOLD (Ray-based Task-Oriented Latent Dynamics) is a new hybrid control architecture for autonomous mobile robots navigating dense, dynamic crowds. Proposed by Seungho Han, Seokju Lee, and Jeonguk Kang, it bridges the gap between purely reactive planning (like Model Predictive Path Integral, MPPI) which suffers from local minima, and long-horizon reinforcement learning. The system leverages a LiDAR-centric latent dynamics model to encode high-dimensional sensor data into a compact state representation, learning a terminal value function and a policy prior. A key innovation is the policy mixture sampling strategy: it augments MPPI candidate trajectories with those from the learned policy, guiding the planner toward goals while maintaining kinematic feasibility.
Extensive tests in stochastic environments with high-density dynamic obstacles show RAY-TOLD outperforms the MPPI baseline, reducing collision rates significantly. The approach combines short-horizon physics-based rollouts with learned long-horizon intent, demonstrating that this blend enhances navigation reliability and safety in complex scenarios. The paper is 8 pages with 4 figures, submitted to arXiv under cs.RO and cs.AI. This work could have practical implications for warehouse robots, autonomous delivery vehicles, and any robot operating in crowded human environments.
- RAY-TOLD blends MPPI (short-horizon physics) with reinforcement learning (long-horizon foresight).
- Uses a LiDAR-centric latent dynamics model to encode sensor data into compact state representations.
- Policy mixture sampling augments MPPI trajectories with learned policy trajectories, reducing collision rates in dense crowds.
Why It Matters
Enables safer, more reliable robot navigation in crowded spaces like warehouses and city sidewalks.