ASTRIL-MPC: Autonomous Traversal Framework of Articulated Tracked Robots with Language-Guided Neural-Kinematic MPC
ASTRIL-MPC is a language-guided neural kinematics model predictive control framework for autonomous traversal of articulated tracked robots in structured, contact-rich environments. It uses a learned kinematics model, NMPC with multi-objective costs and feasibility constraints, and an LLM that proposes bounded updates to weights and bounds through a safety-checked interface. The compiled predictor enables a full control cycle within 100 ms. Across three traversal tasks and a multi-height generalization setting, ASTRIL-MPC improves aggregate traversal-quality score by up to 71% over non-adaptive NMPC and by 67% over a PPO baseline, while eliminating measurable collision impacts during descent.
The framework combines a learned kinematics model with NMPC and an LLM-based parameter adaptation layer. The LLM proposes bounded updates to selected weights and bounds, with range clipping, rate limiting, and consistency checks to ensure safety. The compiled predictor achieves a full control cycle within 100 ms, enabling real-time operation. The approach addresses hybrid and discontinuous robot-terrain interaction by learning short-horizon task-state increments from height sequences and recent trajectories.
This research targets urban search and rescue, where articulated tracked robots must traverse stairwells and cluttered building interiors. The integration of LLMs into low-level control loops suggests a trend toward language-guided autonomy in robotics, potentially reducing the need for hand-tuned controllers and enabling more adaptive behavior in unstructured environments.
Improved traversal quality and collision avoidance can reduce robot damage and mission failure in search and rescue operations, lowering operational costs and increasing reliability. The framework's adaptability may shorten development time for new robot platforms by replacing manual tuning with language-guided parameter updates.
Observable next signals include further validation on physical ATR platforms, extension to more diverse terrains and tasks, and exploration of other LLM interfaces for real-time control. The elimination of collision impacts during descent may lead to adoption in safety-critical applications. The 100 ms control cycle indicates potential for deployment on embedded systems.