Abstract
Legged robots often face environments where vision is unreliable or obstructed. We present PROBE (Proprioceptive Obstacle Detection and Estimation), a framework that allows a quadruped robot to reconstruct a 2D map of its surroundings using only internal sensors. By leveraging a Transformer-based architecture, the robot estimates obstacle geometry and mobility, even detecting hidden objects through indirect contact.
Introduction
In autonomous robotics, vision-based sensors like LiDAR and RGB-D cameras are standard for navigation. However, these sensors struggle in "degraded" conditions such as smoke-filled rooms, low-light environments, or when obstacles are occluded. To address these challenges, researchers from Rutgers University developed PROBE, a system that treats the robot's entire body as a haptic sensor. This "blind" navigation approach allows robots to "feel" their way through cluttered spaces, identifying not just what they are touching, but what those objects are touching in turn.
Methodology: The Obstacle Reconstruction Module (ORM)
The core of PROBE is a hierarchical control and perception framework. At its heart lies the Transformer-based Obstacle Reconstruction Module (ORM).
The ORM utilizes a causal Transformer encoder followed by an MLP decoder. It processes a temporal history of the robot’s internal states, including joint positions, velocities, applied torques, and the robot's 2D pose. The output is a parameterized SE(2) representation of obstacles—defining their position, orientation, dimensions, and mobility (static vs. movable). A key innovation is the system's ability to handle "nested interactions," where the robot pushes a movable object that then contacts a hidden static wall. By analyzing the resulting torque signatures, the ORM can infer the existence of obstacles it has never directly touched.
Experimental Setup and Evaluation
The system was developed and trained within the Isaac Gym simulator, collecting over 180,000 trajectories across varying difficulty levels. These environments ranged from simple single-obstacle setups to complex "Hard" scenarios involving multiple nested obstacles.
For real-world validation, a Unitree Go1 quadruped was deployed in a 4m x 2m corridor. While the robot navigated solely using proprioception, ArUco tags were placed on obstacles to provide ground truth data for performance evaluation.
Results and Key Insights
The study yielded several critical findings:
- Importance of Torque: Ablation studies confirmed that joint torque data is the most vital input for distinguishing between heavy movable objects and fixed structures.
- Reconstruction Accuracy: In simulation, the ORM achieved an average Intersection over Union (IoU) of nearly 0.5 for movable objects. In real-world trials, IoU scores ranged from 0.2 to 0.45, which, despite environmental noise like friction, remained sufficient for successful navigation.
- Hidden Obstacle Detection: The robot successfully hypothesized the presence of static obstacles hidden behind movable boxes, demonstrating a capability that exceeds traditional line-of-sight vision systems.
Conclusion and Future Directions
PROBE demonstrates that proprioception is a powerful, often underutilized modality for robot navigation in challenging environments. While the current implementation assumes box-shaped obstacles and does not explicitly model mass or friction, it provides a robust foundation for "blind" navigation. Future work will explore more complex geometries and the integration of additional non-visual sensors, such as tactile skins, to further enhance environmental awareness.
Summary updated