Source-linked AI summary
When Black-Box Robot Policies Fail: Online Sensor Diagnosis and Repair with a Transparent Surrogate
Marc Ziegler, Batu Özen, Sami Mentes, Thomas Waschulzik, Alois Knoll
TL;DR
When a locomotion policy is frozen or inaccessible, sensor faults pose a monitoring challenge because external monitors cannot inspect its internals. The paper uses a structured surrogate to diagnose and repair inputs while the frozen policy continues to act. Offline MBD-guided repair walks in 429/640 episodes, and online MBD + PRR raises walking from 259/640 to 382/640.
Problem
Method
A structured surrogate separates periodic gait from sensor-dependent corrections, ranks abnormal input contributions, and guides input repair without replacing the frozen policy.
Source: §1 · p. 1Results
Offline MBD-guided repair walks in 429/640 episodes, while online MBD + PRR raises walking from 259/640 to 382/640.
Source: §3 · pp. 3–4Takeaways & Limitations
Clean-action accuracy alone is insufficient evidence of diagnostic value, and the surrogate’s traceable contributions do not reveal the frozen policy’s internal mechanism.
Source: §4 · p. 4Evidence is limited to one policy in simulation, a partly screened fault suite, separately fitted robot surrogates, and simulator-derived references.
Source: §4 · p. 4
Abstract
from arXiv · showhide
Robot foundation models may generalize across tasks and embodiments, but their actions still depend on sensors that can fail. When the policy is frozen or inaccessible, can an external monitor diagnose and repair corrupted inputs without access to its internals? We test a structured surrogate that separates periodic gait from sensor-dependent corrections to guide input repair around a frozen locomotion policy. On simulated Go2, offline diagnosis raises walking success from 259/640 to 429/640 episodes; online repair reaches 382/640, with no triggers in 64 held-out fault-free episodes. A matched boosted-tree surrogate predicts clean actions similarly well but yields no detected repair improvement over no repair. Separate fits improve walking on Go1 and ANYmal C, although phase-aware gradients perform better. Evidence remains limited to one policy and a partly screened simulated fault suite; robot foundation model and hardware validation remain open.
1 Introduction
The paper asks how to monitor sensor-dependent actions when a locomotion policy is inaccessible, since robot foundation models still rely on reliable observations. It introduces a structured surrogate for diagnosis and repair while leaving action generation to the frozen policy.
1 Introduction
- Robot foundation models may transfer across tasks and embodiments [2], but their actions depend on reliable sensor observations and inaccessible policies limit what monitors can inspect.Unusual sensor readings may also reflect legitimate disturbances.Source: §1 · p. 1
- The surrogate separates periodic gait from sensor-dependent corrections, ranks abnormal input contributions, and guides repair without replacing the frozen policy.The frozen policy produces every action.Source: §1 · p. 1
- The work contributes diagnosis and repair without policy gradients, paired offline and online evaluations, and tests on three quadrupeds; direct robot-foundation-model validation remains open.The setting motivates external supervision for robot foundation models.Source: §1 · p. 1
2 Method
The method uses a structured action surrogate to identify sensor contributions associated with abnormal corrections, then applies phase-referenced repair before the frozen policy acts. Diagnosis and repair are evaluated offline and online, with runtime detection calibrated using fault-free episodes.
2 Method
- The surrogate predicts 12 joint-action targets and decomposes corrections into a phase-dependent baseline and individual standardized-input contributions.It represents periodic gait with four Fourier harmonics; clean state–action pairs and alternating policy/surrogate-control targets train the fits.Source: §2 · pp. 1–2
- The phase clock uses simulator-derived contacts for offline fitting and healthy references, but potentially corrupted policy-visible contacts during diagnosis and repair.Touchdown intervals of 8–60 steps update the period estimate.Source: §2 · p. 2
- MBD flags corrections deviating from phase-matched references by more than two standard deviations, then scores sensor units by their contributions to flagged corrections.Offline diagnosis aggregates scores across faulty trajectories and abstains when no corrections are flagged.Source: §2 · p. 2
- Baselines include phase deviation, policy-internal phase-aware gradients, and a matched boosted-tree surrogate with TreeSHAP attribution; all share sensor groups, seeds, repair episodes, and PRR.The boosted-tree comparison uses the same inputs, splits, normalization, nominal gait, residual target, and correction bound.Source: §2 · p. 3
- PRR changes only the selected sensor unit, using a phase-matched healthy value for phase-varying sensors or estimating and subtracting bias otherwise.Contacts and contact ages receive constant nominal values, and repair rules were fixed before evaluation.Source: §2 · p. 3
- Online MBD begins after 25 steps and triggers from calibrated abnormal-correction rates; repair starts on the next step and continues to episode end.Calibration uses 128 fault-free episodes, and runtime requires no policy weights, gradients, or additional policy calls.Source: §2 · p. 3
3 Experiments
In the selected repair suite, MBD improved Go2 walking, and online repair also outperformed no repair. Matched GBM–TreeSHAP did not improve over no repair; separately fitted Go1 and ANYmal C models improved walking, but phase-aware gradients performed better.
Protocol
- Evaluation used 20-second MuJoCo Go2 episodes at 50 Hz and 0.5 m/s, with 640 paired repair episodes per method across 20 main faults, with and without pushes.Ten main faults were selected during development based on preliminary MBD performance; the constructed fault suite is not a sample of hardware failure frequencies.Source: §3 · p. 3
Diagnosis and repair
- MBD-guided repair walked in 429/640 episodes versus 408/640 for phase-aware gradients; the overall gain was 26.6 percentage points, conditional on suite selection.Screened faults rose from 92/288 to 219/288, compared with 167/352 to 210/352 for the other eleven faults.Source: §3 · pp. 3–4, Table 1 · p. 4
- Despite similar clean-action R2 (0.9764 versus 0.9766), GBM–TreeSHAP identified 5/37 faults and walked in 277/640 episodes, not significantly above no repair.Its repair outcome was below MBD (p = 1.2 × 10−33); reused faults and differing masks limit attribution-specific conclusions.Source: §3 · p. 4
Online operation and anchor-contact faults
- Online MBD + PRR raised walking from 259/640 to 382/640, triggered in 0/64 held-out fault-free episodes, and triggered in 454/640 faulty episodes.The median trigger was at step 274; 444 of 454 triggers occurred at the earliest permitted decision.Source: §3 · p. 4
Other embodiments
- Separate Go1 and ANYmal C fits diagnosed 6/13 and 8/13 faults and improved walking on both, while phase-aware gradients reached 602/704 versus MBD’s 570/704.The Go1 and ANYmal C fits were separate; Go2’s method ranking did not transfer.Source: §3 · p. 4, Table 1 · p. 4
4 Discussion and Conclusion
The structured decomposition makes sensor contributions traceable without claiming to reveal the frozen policy’s internal mechanism. Evidence is limited to one policy in simulation and a partly screened fault suite.
4 Discussion and Conclusion
- MBD is the central diagnostic contribution, while PRR provides a controlled test of whether its diagnoses support effective intervention.The surrogate’s traceable sensor contributions do not establish the frozen policy’s internal mechanism.Source: §4 · p. 4
- A matched boosted-tree surrogate achieved similar clean-action accuracy but did not provide comparable diagnostic or repair performance.The paper concludes that clean-action accuracy alone is insufficient evidence of diagnostic value.Source: §4 · p. 4
- Evidence is limited to one policy in simulation, a partly screened fault suite, separately fitted robot surrogates, and simulator-derived references.Future work includes independently specified faults, cross-robot transfer, robust phase estimation, abstention, simultaneous faults, and onboard latency.Source: §4 · p. 4