Teleoperator-Aware and Safety-Critical Adaptive Nonlinear MPC for Shared Autonomy in Obstacle Avoidance of Legged Robots

Ruturaj Sambhus, Muneeb Ahmad, Basit Muhammad Imran, Sujith Vijayan, Dylan P. Losey, Kaveh Akbari HamedarXiv 2025 · DOI · arXiv

TL;DR

This paper introduces a novel adaptive nonlinear Model Predictive Control (MPC) framework for shared autonomy in legged robots, enabling safe and effective human-robot collaboration during obstacle avoidance. By modeling human intent with a noisily rational Boltzmann model and incorporating control barrier functions (CBFs) for safety, the system learns human preferences online while guaranteeing collision-free navigation.

What Problem Does This Paper Address?

Legged robots are increasingly deployed in complex, safety-critical environments like disaster response. While shared autonomy—blending human teleoperation with robot autonomy—promises to leverage human intuition and robotic precision, it faces significant challenges. Human inputs are often noisy and unpredictable, and legged robots themselves are inherently unstable systems with complex dynamics. Existing shared-control methods often use fixed blending strategies that can compromise safety, especially when navigating cluttered environments. This paper tackles the problem of achieving safe and effective shared autonomy for quadrupedal robots, specifically addressing the need to predict and adapt to human intent while ensuring robust obstacle avoidance.

Methods

The core of the proposed solution is a Teleoperator-Aware, Safety-Critical, Adaptive Nonlinear MPC (ANMPC) framework. This framework operates hierarchically:

  • High-Level CBF-based ANMPC (10 Hz): This layer generates blended human-robot velocity references. It uses a noisily rational Boltzmann model to predict human intent, with parameters adapted online using a Projected Gradient Descent (PGD) law based on observed joystick commands. Crucially, it incorporates Control Barrier Functions (CBFs) to enforce safety constraints over a prediction horizon, ensuring forward invariance of safe sets despite human uncertainty.
  • Mid-Level Dynamics-Aware NMPC (60 Hz): This layer enforces reduced-order Single Rigid Body (SRB) dynamics to track the velocity references from the high-level controller. It optimizes ground reaction forces to ensure dynamic stability and satisfy friction cone constraints.
  • Low-Level Nonlinear Whole-Body Controller (500 Hz): This layer imposes the full-order robot dynamics via quadratic programming (QP) to track the mid-level trajectories.

A key innovation is the integration of human intent modeling within the MPC framework. Instead of directly solving for the human action, the system enforces an algebraic constraint derived from the human's value function, allowing for efficient computation. The human's value function is parameterized to encourage goal-reaching, penalize human effort, and enforce obstacle avoidance.

Datasets

Not discussed in the paper. The experiments utilize a Unitree Go2 quadrupedal robot in a controlled laboratory environment with cylindrical obstacles and designated target goals.

Key Insights and Ideas

  • Adaptive Human Intent Modeling for Safety: The framework's ability to learn human intent parameters online using a PGD law is crucial. This allows the robot to adapt to individual user behaviors and preferences, leading to more intuitive and effective collaboration.
  • CBFs for Guaranteed Safety in Shared Autonomy: Integrating CBFs into the MPC framework provides formal guarantees of safety, ensuring that the robot avoids collisions even with unpredictable human inputs. This is a significant advancement over traditional safety methods.
  • Hierarchical Control for Efficiency and Performance: The layered approach (high-level intent/safety, mid-level dynamics, low-level whole-body control) effectively balances computational complexity with performance. It allows for real-time execution on hardware while leveraging sophisticated control strategies.
  • Bridging the Gap Between Human Intuition and Robotic Precision: The fixed arbitration weight, combined with adaptive intent modeling, strikes a balance between allowing human control and ensuring the robot's autonomous capabilities are utilized for safety and efficiency.

Key Results

  • Real-time Obstacle Avoidance: The framework successfully enabled real-time obstacle avoidance during extensive hardware experiments and user studies on a Unitree Go2 robot.
  • Online Adaptation of Human Intent: The PGD-based adaptation law effectively updated human intent parameters online, as demonstrated by the time evolution plots.
  • Safe Teleoperator Collaboration: User studies indicated that the shared autonomy approach facilitated safe and effective collaboration, with participants finding the assistive behaviors perceptible and helpful.
  • Improved Safety Margins: The CBF-based ANMPC maintained larger safety margins compared to a baseline fully human-controlled approach.
  • Computational Efficiency: The NLP solver achieved an average solve time of 27 ms, enabling real-time operation.

Limitations and Future Work

  • User Perception of Control: User studies revealed that participants sometimes felt more comfortable with baseline control, possibly due to not being informed about the active control mode, leading to occasional conflicts between user intent and assistance.
  • IMU Drift: IMU drift occasionally caused minor collisions when the robot traveled longer distances, suggesting a need for improved state estimation.
  • Suboptimal Trajectories from Novices: Novice users sometimes generated suboptimal trajectories that conflicted with the ANMPC predictions, highlighting the importance of user training and transparency.

Future work suggested includes:

  • Reducing drift through better state estimation.
  • Enhancing user collaboration through training.
  • Investigating more sophisticated methods for transparency and user understanding of the autonomy.

Summary updated

BibTeX
@misc{sambhus2025teleoperator,
  author = {Ruturaj Sambhus and Muneeb Ahmad and Basit Muhammad Imran and Sujith Vijayan and Dylan P. Losey and Kaveh Akbari Hamed},
  title = {{Teleoperator-Aware and Safety-Critical Adaptive Nonlinear MPC for Shared Autonomy in Obstacle Avoidance of Legged Robots}},
  year = {2025},
  eprint = {2509.22815},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.22815},
}
Figure