Adaptive Reinforcement and Model Predictive Control Switching for Safe Human-Robot Cooperative Navigation

TL;DR

This work tackles safe, human-guided mobile navigation under proximity and collision constraints by integrating a PPO-trained follower with a one-step MPC safety filter through an adaptive neural switcher, aided by a decoupled perception stack (LiDAR-based latent encoding and temporal human-motion encoding). The approach, ARMS, uses soft action fusion with an EMA-smoothed weight to balance conservative safety enforcement and agile learned navigation, achieving high success rates in cluttered environments while dramatically reducing computational latency compared to multi-step MPC baselines. Key contributions include the learned adaptive arbitration mechanism, the decoupled perception scheme, and extensive simulations, Gazebo tests, and real-world deployment demonstrating practical robustness and real-time performance. The results suggest ARMS provides a scalable, real-time solution for safe human-robot collaboration in dynamic, partially observed environments, with potential extensions to moving obstacles and multi-human scenarios, and a publicly available implementation at the project repository.

Abstract

This paper addresses the challenge of human-guided navigation for mobile collaborative robots under simultaneous proximity regulation and safety constraints. We introduce Adaptive Reinforcement and Model Predictive Control Switching (ARMS), a hybrid learning-control framework that integrates a reinforcement learning follower trained with Proximal Policy Optimization (PPO) and an analytical one-step Model Predictive Control (MPC) formulated as a quadratic program safety filter. To enable robust perception under partial observability and non-stationary human motion, ARMS employs a decoupled sensing architecture with a Long Short-Term Memory (LSTM) temporal encoder for the human-robot relative state and a spatial encoder for 360-degree LiDAR scans. The core contribution is a learned adaptive neural switcher that performs context-aware soft action fusion between the two controllers, favoring conservative, constraint-aware QP-based control in low-risk regions while progressively shifting control authority to the learned follower in highly cluttered or constrained scenarios where maneuverability is critical, and reverting to the follower action when the QP becomes infeasible. Extensive evaluations against Pure Pursuit, Dynamic Window Approach (DWA), and an RL-only baseline demonstrate that ARMS achieves an 82.5 percent success rate in highly cluttered environments, outperforming DWA and RL-only approaches by 7.1 percent and 3.1 percent, respectively, while reducing average computational latency by 33 percent to 5.2 milliseconds compared to a multi-step MPC baseline. Additional simulation transfer in Gazebo and initial real-world deployment results further indicate the practicality and robustness of ARMS for safe and efficient human-robot collaboration. Source code and a demonstration video are available at https://github.com/21ning/ARMS.git.

Figures (6)

• Figure 1: Human-guided collaborative transport. The cobot executes real-time local navigation while carrying a payload, and the human provides high-level intent through motion. Proximity and safety constraints must be satisfied simultaneously in shared workspaces.
• Figure 2: Overview of the proposed ARMS framework. A learned follower (trained via PPO) and a one-step MPC-formulated safety filter independently generate velocity commands from heterogeneous sensory inputs. A learned switcher observes compact risk features and softly fuses the two actions using an EMA-smoothed gating signal. When the safety QP is infeasible, execution falls back to the follower action.
• Figure 3: Ablation analysis of Follower variants in Scenario 3, averaged over a total of 300 evaluation trials. (a) Success rate across distinct navigation scenarios. (b) Average obstacle clearance maintained by the corresponding configurations.
• Figure 4: Comparative ablation of the switcher architecture in Scenario 3, illustrating the effect of feature selection and fusion strategies on navigation success. Results are averaged over a total of 300 evaluation trials, with shaded bands and error bars indicating the 95% confidence intervals.
• Figure 5: Representative Gazebo evaluation in a cluttered hallway. Two static obstacles are placed along the corridor while a human actor follows a goal-directed trajectory. The robot receives 2D LiDAR scans and outputs velocity commands. The bottom row shows the corresponding RViz visualizations: the red circle denotes the desired following region around the operator, and the red curve indicates the operator trajectory. The robot stays within the following region while maneuvering around obstacles and maintaining a safe separation distance.