Standing Tall: Sim to Real Fall Classification and Lead Time Prediction for Bipedal Robots

TL;DR

This work extends a nominal fall prediction algorithm to real-time online operation on the full-sized bipedal robot Digit, validating performance in both simulation and hardware. It demonstrates comparable online and offline results, with a recovery rate of $0.97$ in simulation and a mean predicted lead time around $1.1$s, well above the required $0.2$s, and a zero false-positive rate in simulation. The study further analyzes robustness to omnidirectional faults and introduces a minimal-data fine-tuning strategy that substantially improves all metrics, including reducing the average lead-time error and false positives. The results indicate strong sim-to-real transfer, highlight certain limitations under frontal-plane disturbances, and provide a practical path toward robust real-time fall prediction for standing tasks in bipedal robots.

Abstract

This paper extends a previously proposed fall prediction algorithm to a real-time (online) setting, with implementations in both hardware and simulation. The system is validated on the full-sized bipedal robot Digit, where the real-time version achieves performance comparable to the offline implementation while maintaining a zero false positive rate, an average lead time (defined as the difference between the true and predicted fall time) of 1.1s (well above the required minimum of 0.2s), and a maximum lead time error of just 0.03s. It also achieves a high recovery rate of 0.97, demonstrating its effectiveness in real-world deployment. In addition to the real-time implementation, this work identifies key limitations of the original algorithm, particularly under omnidirectional faults, and introduces a fine-tuned strategy to improve robustness. The enhanced algorithm shows measurable improvements across all evaluated metrics, including a 0.05 reduction in average false positive rate and a 1.19s decrease in the maximum error of the average predicted lead time.

Figures (8)

• Figure 1: Real-time detection of a critical fault introduced between the first and second panels.
• Figure 2: The nominal fall prediction algorithm is comprised of three components: a critical fault classifier, a lead time classifier, and a lead time regressor. The green boxes depict the algorithm's predicted lead time.
• Figure 3: Kinematics architecture of the Digit robot by Agility Robotics ref:ar_digitref:mungai2024fall.
• Figure 4: Showcases the results of switching vs not switching to a recovery controller when the nominal fall prediction algorithm correctly detects a critical abrupt fault. The fault is introduced at 15s.
• Figure 5: The nominal fall prediction algorithm detects a critical incipient fault and switches to the recovery algorithm