Language-Guided Terrain-Adaptive Neural MPC for Autonomous Traversal of Articulated Tracked Robots

A terrain-conditioned neural kinematics model fuses terrain and recent motion history to capture hybrid-contact effects without explicit contact-mode enumeration. Relative to classical methods that rely on pre-defined configurations or explicit contact modeling, this model takes a map-attached local height sequence and a length-k history buffer as input and directly predicts short-horizon task-state increments, bypassing the difficulty of analytically modeling hybrid contacts. The paper defines the model inputs and outputs (Eqs. 7a–7b) and a control-aware composite loss with fit, overshoot, smoothness, magnitude, and regularization terms (Table I), and states that the compiled predictor supports a full control cycle within 100 ms.

The learned kinematics is embedded in NMPC as the state-transition constraint, solving a finite-horizon optimal control problem with explicit state and input constraints. Compared with pure feedforward policies or reinforcement learning, this design retains hard feasibility constraints and a tunable multi-objective cost structure while replacing hard-to-solve contact dynamics with a data-driven model. The paper provides the full optimization problem (Eqs. 11a–11f), including initial-state clamping, input box constraints, and state bounds, and reports solving the full problem within 100 ms.

An LLM serves as a terrain-adaptation regulator, translating natural-language guidance and recent terrain-dependent behavior into bounded edits of cost weights, input bounds, and flipper PD gains. Rather than generating actions directly, the LLM acts as a high-level optimizer over a restricted parameter set, with a safety gate enforcing editable-set restrictions, clipping, rate limits, smoothing, and feasibility checks, enabling fast terrain adaptation without altering the MPC formulation or bypassing its feasibility constraints. The paper presents the prompt construction and parameter extraction pipeline (Eqs. 14a–14b), the guarded update formulas (Eqs. 15a–15c), and the parameter categories adapted online (Table II), and reports that this mechanism improves traversal quality by 67% over PPO and enables terrain-dependent generalization without retraining the predictor or changing the solver.

The framework is validated across three traversal tasks, a multi-height generalization setting, and real-robot trials over four indoor obstacles. Compared with non-adaptive neural MPC and a PPO baseline, the framework improves an aggregate traversal-quality score by up to 71% and 67% respectively, eliminates measurable collision impacts during descent, and real-robot trials further demonstrate transfer to contact-rich physical traversal. Evidence comes from the paper's reported simulation task comparisons and real-robot trials; the abstract gives relative improvement magnitudes and the qualitative result of eliminating collision impacts, but does not provide sample sizes or statistical significance details at the abstract level.

This work targets autonomous traversal of articulated tracked robots in structured but contact-rich environments, restricts the state to the 2D sagittal plane, and serves as a downstream module tracking predefined trajectories rather than handling global path planning; its design suits scenarios requiring hard feasibility constraints and fast terrain adaptation, and physical transfer has been demonstrated in real-robot trials over four indoor obstacles.

At the abstract level, the paper does not provide sample sizes, statistical significance, error bars, or ablation details, so the robustness of the relative improvements still needs to be verified in the main text; the trigger frequency of LLM-guided adaptation, sensitivity to prompt design, and generalization boundaries across broader terrains and tasks are also directions a careful reader may continue to watch.