Updating Robot Safety Representations Online from Natural Language Feedback

Leonardo Santos, Zirui Li, Lasse Peters, Somil Bansal, Andrea Bajcsy

Introduction

Traditional robotic safety frameworks typically rely on static environmental maps and pre-calculated safety controllers. While effective for avoiding permanent structures like walls, these "baked-in" safety measures often fail to account for the dynamic and semantic nature of human-centric environments. Real-world scenarios involve transient hazards—such as a coffee spill, a fragile rug, or a cordoned-off area—that are not physical barriers but must be avoided to ensure operational safety and respect user preferences.

The challenge in addressing these semantic constraints is twofold: first, the robot must perceive and interpret high-level human intent (e.g., "Avoid the workout zone"); second, it must update its mathematically rigorous safety controllers in real-time without the prohibitive computational cost of re-calculating safety sets from scratch.

Methodology: The Safe-Lang Framework

To bridge the gap between high-level language and low-level control, we propose Safe-Lang, a pipeline that integrates Vision-Language Models (VLMs) with Hamilton-Jacobi (HJ) reachability analysis.

The Safe-Lang system architecture, showing the transition from offline safety priors to online updates driven by human language and VLM-based perception.

Perception and Semantic Mapping

When a human provides a natural language command, the system employs a pre-trained VLM (OWLv2) to identify the target objects in the robot's RGB-D camera feed. These detections are projected onto a 2D ground plane to define a "Semantic Failure Set"—a spatial representation of regions the robot must avoid.

Formal Safety via HJ Reachability

Safe-Lang utilizes HJ reachability to calculate a Backward Reachable Tube (BRT). The BRT identifies all states from which a robot is guaranteed to eventually enter the failure set, regardless of its control inputs. This provides a formal "safety shield" that monitors the robot's state relative to "doomed" configurations.

Efficient Online Updates

Calculating BRTs is traditionally computationally intensive. Safe-Lang overcomes this by using a warm-starting technique. Instead of re-calculating safety values from a blank slate, the system initializes the HJ reachability computation with previously known safety values. Since semantic updates are typically local, this allows the safety representation to be updated in approximately 3 seconds, enabling online deployment.

Experimental Evaluation

The framework was evaluated in both high-fidelity simulations and real-world hardware deployments.

  • Simulation: Using the Habitat 3.0 simulator with the HSSD-HAB dataset, the robot was tasked with navigating "Home Gym" and "Hallway" environments while respecting constraints like "avoid the expensive rug."
  • Hardware: A LoCoBot mobile manipulator equipped with an Intel RealSense camera was used to validate the system's ability to avoid real-world hazards like caution tape and spills based on live feedback.

Results and Discussion

Experimental results demonstrate that Safe-Lang is the only method that consistently avoids both physical obstacles and newly defined semantic constraints.

Trajectory comparisons in simulated environments. Safe-Lang (orange) successfully navigates around semantic constraints that traditional planners (black/green) or unprotected language-based planners (purple) fail to avoid safely.

Key findings include:

  • Robust Safety Filtering: The safety filter acts as a policy-agnostic shield, intervening only when the robot's nominal planner (e.g., MPPI) attempts to enter a "doomed" state.
  • Language Flexibility: VLMs effectively map subjective human labels to physical coordinates, allowing the robot to respect constraints that are invisible to traditional sensors like LiDAR.
  • Computational Viability: The warm-starting approach reduced update times to a level suitable for real-time interaction, maintaining safety even when commands were issued shortly before a potential collision.

Conclusion and Future Work

Safe-Lang provides a robust mechanism for updating robot safety representations online via natural language. While the current implementation uses rectangular bounding boxes, future iterations could leverage pixel-level semantic segmentation for more precise boundaries. Additionally, incorporating uncertainty-aware perception and handling dynamic semantic constraints (e.g., moving pets) remain promising avenues for future research.

Summary updated

BibTeX
@inproceedings{santos2025updating,
  author = {Leonardo Santos and Zirui Li and Lasse Peters and Somil Bansal and Andrea Bajcsy},
  title = {{Updating Robot Safety Representations Online from Natural Language Feedback}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2025},
  doi = {10.1109/icra55743.2025.11127680},
}
Figure