TL;DR
This paper introduces a "radar-native" safety filter that uses direct Doppler measurements from FMCW radar as the derivative for Control Barrier Functions (CBFs). By bypassing the need for noisy and delayed obstacle-state estimators, the system achieves significantly faster reaction times and fewer safety violations in dynamic environments.
What Problem Does This Paper Address?
In robotics, ensuring a safe distance from moving obstacles is a fundamental challenge. Control Barrier Functions (CBFs) are a popular mathematical tool for this, but they require knowing the "range rate"—how fast the distance between the robot and an obstacle is changing.
Most sensors, like LiDAR or depth cameras, only measure distance. To find the range rate, robots must either differentiate successive measurements (which is noisy) or use complex estimators like Kalman filters (which introduce delay). This "perception gap" is dangerous: if an obstacle suddenly accelerates toward a robot, the estimator might take several frames to catch up, leading to a collision or a "buffer incursion" where the robot gets closer than its defined safety margin.
Methods
The authors propose a formulation that treats Frequency-Modulated Continuous-Wave (FMCW) radar as a "native" safety sensor. Unlike LiDAR, FMCW radar measures the radial relative velocity of every detection directly via the Doppler effect.
Key technical contributions include:
- Doppler as the Barrier Derivative: The measured Doppler range rate is used directly as the derivative ($\dot{h}$) of the barrier function. This eliminates the need for an external obstacle-velocity estimator.
- Closed-Form Safety Filter: The authors derive a closed-form solution for the robot's forward velocity command. This is computationally efficient because it doesn't require solving a Quadratic Program (QP) at every timestep.
- Measurement-Robust Margin: They define a safety margin ($\epsilon_{meas}$) that accounts for bounded errors in range, Doppler, ego-velocity, and bearing. This margin is based on verifiable sensor specifications rather than arbitrary assumptions about obstacle behavior.
Datasets
The researchers evaluated their approach using:
- Paired Simulation Sweep: A set of 560 simulated encounters (508 evaluable) with varying initial ranges (0.6m to 2.0m) and approach speeds (0.2m/s to 1.4m/s). This allowed for a direct "head-to-head" comparison between Doppler-based control and traditional Kalman filter-based methods.
- Hardware Experiments: Tests performed on a Unitree Go1 quadruped robot equipped with a Texas Instruments IWR6843AOPEVM FMCW radar.
Key Insights and Ideas
- Zero-Latency Response: Because Doppler is measured within a single radar frame, the safety filter reacts to changes in obstacle motion almost instantly. In simulations, the Doppler-based command changed 0.09s after an obstacle started moving, whereas a constant-velocity Kalman filter took 0.33s to respond.
- Handling Multi-Object Ambiguity: A robot might be close to a stationary wall but have a person running toward it from further away. Traditional "nearest-neighbor" filters might focus on the wall. The radar-native CBF correctly identifies that the farther, faster-approaching object is the greater threat and prioritizes braking for it.
- Verifiable Safety: The safety guarantees in this paper are tied to the physical properties of the sensor (like range bin width and Doppler resolution). This moves safety from "hopeful estimation" toward "verifiable engineering."
Key Results
- Safety Reliability: In 508 paired simulation encounters, the direct Doppler method resulted in only 2 buffer incursions, compared to 15 for a constant-acceleration Kalman filter and 25 for a constant-velocity Kalman filter.
- Predictable Standoff: In hardware tests, the robot consistently stopped at a mean distance of 0.777 ± 0.015 m, closely matching the theoretically predicted stopping distance of 0.758 m.
- Robustness to Noise: Even when the radar's detection probability was reduced to 60% in simulations, the Doppler-based controller remained significantly safer than range-only estimators.
Limitations and Future Work
- Longitudinal Only: The current formulation only certifies safety for the longitudinal (forward/backward) channel. It does not yet account for lateral safety or coupling the range constraint with steering commands.
- Continuous Detections: The theoretical safety guarantee assumes that detections are continuously available. While the system handles some missed detections in practice, "ghost" detections or total sensor occlusion remain challenges.
- Future Directions: The authors aim to extend this work to include sampled-data updates, cover tangential motion, and integrate steering into the safety manifold.
Summary updated