PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Zhenfeng Gan, Yanbo Chen, Lirong Che, Yongyi Ma, Rongkai Zhu, Xueqian Wang

arXiv:2609.13083v3cs.ROcs.AI

TL;DR

Contact-rich traversal by articulated tracked robots is difficult to model, and existing control approaches may lack explicit optimization over future contacts and constraints. The paper combines terrain-conditioned neural kinematics with constrained MPC and guarded LLM parameter updates; its reported overall mean improves over fixed-parameter neural MPC and PPO.

  • Problem

    Contact-rich robot–terrain interactions are hybrid and discontinuous, making accurate ATR modeling difficult, while feedforward action policies do not explicitly optimize future contacts or task-specific constraints.

    Source: §I · p. 1

  • Method

    The framework embeds terrain-conditioned neural kinematics in constrained NMPC and uses a safety-gated LLM to make bounded updates to selected controller parameters.

    Source: §I · p. 2

  • Results

    The overall mean improves by 42% over fixed-parameter neural MPC and by 67% over PPO.

    Source: 3) Results · p. 6

  • Takeaways & Limitations

    Experiments indicate that terrain-conditioned prediction and language-guided adaptation can support real-time, feasible ATR traversal in structured contact-rich environments.

    Source: §V · p. 7

Abstract

from arXiv · show

In urban search and rescue, articulated tracked robots (ATRs) must traverse structured but contact-rich environments such as stairwells and cluttered building interiors. Reliable autonomy remains challenging because robot-terrain interaction (RTI) is hybrid and discontinuous, and effective flipper-track coordination is difficult to model analytically. We present ASTRIL-MPC, a language-guided neural kinematics model predictive control (MPC) framework for autonomous traversal. A learned kinematics model predicts short-horizon task-state increments from a height sequence and recent trajectories; NMPC plans with multi-objective costs and strict feasibility constraints; and a large language model (LLM) proposes bounded updates to selected weights and bounds through a safety-checked interface with range clipping, rate limiting, and consistency checks. The compiled predictor enables a full control cycle within 100 ms. Across three traversal tasks and a multi-height generalization setting, ASTRIL-MPC improves an aggregate traversal-quality score by up to 71% over a non-adaptive NMPC and by 67% over a PPO baseline, while eliminating measurable collision impacts during descent. These results indicate that combining terrain-conditioned neural kinematics, optimization-based planning, and language-guided adaptation yields data-efficient and robust autonomy for articulated tracked robots. Real-robot trials over four indoor obstacles further demonstrate transfer to contact-rich physical traversal.

I. INTRODUCTION

Articulated tracked robots face difficult-to-model, discontinuous terrain contacts and need coordinated flipper-track control. The paper combines terrain-conditioned neural prediction, constrained MPC, and guarded language-guided parameter adaptation; related work highlights limits in geometric, RL, and direct language-to-action approaches.

I. INTRODUCTION

  • Hybrid, discontinuous robot–terrain interactions make explicit modeling and coordinated chassis–flipper motion difficult, while terrain geometry can change the preferred stability, impact, and speed trade-off.The motivating setting is urban search and rescue, including stairwells and cluttered interiors.Source: §I · p. 1
  • MPC provides hard constraints, tunable multi-objective costs, and receding-horizon planning, but ATR deployment is hindered by contact-modeling difficulty and expert-dependent scenario tuning.The proposed framework pairs those MPC properties with learned kinematics and language-guided tuning.Source: §I · p. 1
  • The LLM acts as a high-level optimizer that makes bounded edits to MPC weights and bounds, while the low-level neural MPC retains hard feasibility constraints and closes the control loop.The safety gate applies editable-set restrictions, clipping, rate limits, smoothing, and feasibility checks.Source: §I · p. 2
  • A terrain-conditioned neural kinematics model fuses terrain and recent motion history to predict hybrid-contact effects without explicit contact-mode enumeration, providing a lightweight basis for planning.The model is embedded in NMPC as its state-transition constraint, with explicit state and input constraints and a full solve within 100 ms.Source: §I · p. 2
  • Terrain-adaptive neural MPC improves traversal quality by 67% over PPO and enables terrain-dependent generalization without retraining the predictor or changing the solver.The LLM updates selected weights, input bounds, and flipper PD gains through a safety gate.Source: §I · p. 2

II. RELATED WORK

  • Geometric and hybrid planners rely on hand-crafted templates or mode switching [8],, –, while flipper-robot RL faces reward-shaping and training-stability challenges,, –.Direct VLA action generation also lacks explicit hard state/input constraints and terrain-dependent optimal control [1], [2].Source: §II · p. 2

III. METHODOLOGY

The architecture combines map-attached neural kinematics, constrained receding-horizon optimization, and guarded LLM parameter updates. Its state representation uses a fixed local height sequence and recent state-input history to describe motion relative to terrain.

A. System Overview

  • A compiled neural kinematics model predicts short-horizon state increments from terrain and recent motion, while NMPC optimizes under multi-objective costs and hard feasibility constraints.The LLM proposes bounded edits to selected weights, bounds, and flipper gains; range, rate, and consistency checks guard updates.Source: A. System Overview · p. 3

1) State Representation:

  • The model uses a sagittal x–z state because explicit contact modeling is too complex for real-time use, while roll stability is left to globally roll-favorable path selection.The system focuses downstream on articulated motion to track predefined trajectories and negotiate obstacles.Source: 1) State Representation · p. 3
  • A map-attached representation describes robot state relative to a terrain profile sampled around the longitudinal COM position and held fixed over the prediction horizon.The state includes local longitudinal offsets anchored at zero; the task vector describes absolute state, while the state vector describes terrain-relative state.Source: 1) State Representation · p. 3
  • A length-k buffer of recent states and control inputs supplies short-term motion context to the learned kinematics and NMPC.Here k is the number of most recent time steps.Source: 1) State Representation · p. 3

2) Neural Kinematics Model:

The neural kinematics model predicts task-space increments using terrain, recent motion, and stage controls. Its composite training loss balances prediction accuracy with smoothness and robustness.

2) Neural Kinematics Model:

  • The learned model predicts each task-space increment from the fixed local terrain sequence, recent motion history, and stage control, then updates the state by that increment.Predicted states use the physical components x, z, pitch, and front and rear flipper angles.Source: 2) Neural Kinematics Model · p. 3
  • The composite loss balances accuracy, smoothness, and robustness using weighted fitting, forward-overshoot penalties, temporal smoothness, plausible-magnitude constraints, and weight regularization.Overshoot penalizes positive forward-direction prediction bias; smoothness aligns predicted forward increments with the local trend.Source: 2) Neural Kinematics Model · p. 3

C. Model Predicted Control

The NMPC uses learned kinematics and the fixed terrain profile to optimize task states and control inputs over a finite horizon. Its formulation combines multi-objective costs with strict feasibility constraints.

C. Model Predicted Control

  • At each control step, NMPC optimizes task-space states and inputs using learned kinematics and the fixed height sequence, subject to multi-objective costs and strict feasibility constraints.The optimization is finite-horizon.Source: C. Model Predicted Control · p. 4
  • Input penalties are centered on a feedforward control sequence aligned to the NMPC grid, with diagonal state and input weights.The state weights include terminal weights.Source: 1) Cost and Reference · p. 4, C. Model Predicted Control · p. 4

1) Cost and Reference:

The input-cost reference includes flipper PD feedforward and a ramped chassis forward-velocity reference. Clipping and rate limiting keep the feedforward sequence feasible and smooth its application.

1) Cost and Reference:

  • Flipper feedforward inputs are generated by per-joint PD laws, and the chassis forward-velocity reference is ramped across the horizon to suppress first-stage jerk.The speed ramp begins from the measured speed.Source: 1) Cost and Reference · p. 4
  • Each feedforward input is clipped to the actuator box and subjected to the applied-control rate limit and smoothing factor before entering the cost.This keeps the penalty center feasible and reduces early-stage optimizer effort.Source: 1) Cost and Reference · p. 4
  • Language guidance and recent terrain-dependent history inform LLM updates to selected MPC parameters, which are applied to costs and constraints and persist until new guidance arrives.The updates include tracking weights, input penalties, flipper gains, smoothing and rate limits, and speed references and bounds.Source: Figure 4 · p. 4

2) Constraints:

The controller keeps inputs within a feasible set and lets an LLM propose restricted edits to selected MPC parameters. Safety checks bound and smooth those edits before application.

2) Constraints:

  • Control inputs are constrained to a bounded feasible set, including explicit forward-speed limits that may be adjusted online under safety gating.The initial state is clamped to the current measurement.Source: 2) Constraints · p. 4

D. Language-Guided Terrain Adaptation via LLM

  • The LLM uses operator guidance and recent terrain-dependent behavior to suggest edits to MPC weights and bounds, rather than generating actions directly.Its prompt includes current measurements, recent history, active parameters, and an optimization template.Source: D. Language-Guided Terrain Adaptation… · p. 4
  • The editable set covers state-tracking weights, input penalties, flipper PD gains, smoothing factors, and velocity bounds; other components remain fixed.During online operation, an instruction triggers a guarded update before the controller rolls out the learned model and solves NMPC.Source: D. Language-Guided Terrain Adaptation… · p. 5
  • A safety gate checks proposed edits against bounds, per-update rate limits, smoothing, and feasibility conditions before applying them.The checks include vmin < vmax and positive-definite W, We, and R; the perception stack, learned model, and solver remain unchanged.Source: D. Language-Guided Terrain Adaptation… · p. 5

A. Experimental Setup

The experiments use a tracked rescue robot in simulated height-map terrain, with a compiled neural kinematics model and a 100 ms NMPC cycle. The predictor is trained on simulated stair-traversal data.

1) Platform and Simulation Setup:

  • The tracked rescue platform has four independently actuated flippers and a symmetric chassis; simulated terrains are height maps sampled at 0.0625 m.The full NMPC loop runs within 100 ms per cycle.Source: 1) Platform and Simulation Setup · p. 5

2) Hardware and Software:

  • The NMPC is modeled in CasADi, the learned kinematic model is compiled via L4CasADi, and ACADOS solves the resulting nonlinear program online.Experiments run on an Intel i5 CPU with an NVIDIA RTX 2080Ti GPU.Source: 2) Hardware and Software · p. 5

3) Default Parameters:

  • Default settings use horizon H=12 and ∆t=0.1 s, with specified state and input weights, flipper PD gains, update limits, smoothing, and speed targets.The reported speed target is 0.30 m/s, with initial command 0.12 m/s.Source: 3) Default Parameters · p. 5

B. Data Collection and Training Details

  • The next-step predictor maps a 64-bin local height sequence, recent positions, commands, and short-horizon states to the 5D task state at t+1.Samples are recorded at 10 Hz.Source: B. Data Collection and Training Details · p. 5
  • Training uses 5k simulated stair-ascent and -descent samples split 95:5, with a residual feedforward MLP and a control-aware loss.The model is trained with Adam at learning rate 1×10^-4 and batch size 2048.Source: B. Data Collection and Training Details · p. 5

C. Evaluation and Analysis of Neural Kinematics

The full loss design produces the most accurate and stable neural-kinematics rollouts across task-state components. The evaluation measures open-loop errors over prediction windows and compares three loss configurations.

C. Evaluation and Analysis of Neural Kinematics

2) Results and Analysis:

  • Type 3, the full loss, yields the most accurate and stable rollouts across all state components, while Type 1 drifts and Type 2 reduces but does not eliminate errors.The trajectory visualization and per-step RMSE curves show consistent trends across variables.Source: Figure 6 · p. 6, Figure 5 · p. 6, 2) Results and Analysis · p. 6

1) Evaluation Protocol and RMSE Aggregation:

  • For each trajectory, the protocol freezes height and history at each non-overlapping H=10 window, then computes per-step RMSE across windows for five task-state components.The components are x, z, pitch, and front and rear flipper angles.Source: 1) Evaluation Protocol and RMSE… · p. 6

D. Comparative Experiment

  • The comparative experiment identifies Proximal Policy Optimization (PPO) as the reinforcement-learning baseline.The supplied passage identifies the baseline but does not report comparative results here.Source: D. Comparative Experiment · p. 6

1) Evaluation Tasks:

Three traversal scenarios assess success, while normalized indicators summarize stability, smoothness, and impact. Reported metric values are means across repeated runs.

1) Evaluation Tasks:

  • The evaluation covers single-block crossing, stair ascent, and stair descent, with three runs per controller and success defined as reaching the goal within 60 seconds without tipping.Each scenario is executed three times per controller.Source: 1) Evaluation Tasks · p. 6

2) Performance Metrics:

  • Six raw indicators—roll and pitch angles, their angular velocities, and forward and vertical accelerations—form normalized stability, smoothness, and impact penalties, with lower values better.Each pair is normalized to [0, 1], and reported values are means across repetitions.Source: 2) Performance Metrics · p. 6

3) Results:

Terrain adaptation improves traversal penalties across terrains, and tests show transfer across obstacle heights and to physical traversal over four indoor obstacles. The conclusion attributes real-time feasibility to compiling the neural kinematics into the solver and identifies broader terrain and energy-aware objectives as future work.

3) Results:

  • 42% improvement in overall mean over fixed-parameter neural MPC and 67% over PPO accompany 33% ascent and 71% descent gains over fixed-parameter neural MPC.Table III compares three normalized penalties and their aggregate mean; adaptation reduces the Shock penalty from 3.0×10−3 to 1.4×10−4, with the largest benefits on dynamic stair scenarios.Source: 3) Results · pp. 6–7, Table III · p. 7

E. Generative Experiment

  • Across multiple obstacle heights, the controller maintains feasible configurations and bounded chassis pitch without task-specific retraining.It coordinates front-flipper probing and rear-flipper bracing; stable bounded LLM updates indicate transfer to unseen geometric conditions.Source: E. Generative Experiment · p. 7

F. Real-Robot Validation

  • Physical trials demonstrate contact-rich traversal across four indoor obstacles with different geometries.Acceleration-magnitude residuals indicate vibration and contact impacts, while IMU-integrated speed is used only as a qualitative motion indicator because it can drift.Source: F. Real-Robot Validation · p. 7, Figure 8 · p. 7

V. CONCLUSION

  • The conclusion reports that compiling neural kinematics into the solver accelerates optimization for real-time control on articulated tracked robots.The experiments also indicate improved multi-step prediction and more stable traversal from the learned kinematics, composite loss, and language-guided terrain adaptation.Source: §V · p. 7
  • Future work targets deformable or variable-geometry objects, energy-aware objectives, additional platforms, and a broader range of terrains and contact conditions.Energy-aware objectives are proposed to improve efficiency in long-duration missions.Source: §V · p. 7

Related papers

Loading 2609.13083v3…