Counterfactual Weight-Space Repair (CWSR): Mitigating Heavy-Tail Autonomous Driving Failures via Null-Space Projected Synapse Surgery

Neural control policies trained via reinforcement learning or imitation learning achieve low error rates across nominal operating distributions (99.9% of routine trajectories), yet exhibit brittle failure modes in rare, high-severity tail regimes (0.1% edge cases such as abrupt tire-road friction drops on black ice mu = 0.14, dense-fog occluded pedestrian emergence, or asymmetric split-mu hydroplaning). Standard remedies—such as dense gradient fine-tuning (SGD/PPO) on synthetic counterexamples—modify 100% of network parameters globally, inducing catastrophic forgetting on nominal highway trajectories and creating conflicting control reflexes across coupled hazard scenarios. In this paper, we formulate Counterfactual Weight-Space Repair (CWSR), a non-gradient mechanistic intervention framework that localizes and patches a sparse subset (<11%) of causally responsible synapses directly in weight space. Using stochastic Monte Carlo trajectory ensembles under coupled nonlinear tire dynamics, we partition rollouts into a surviving safe subset characterized by positive Safety Margin at Risk (SMaR_alpha) and a heavy-tail failure subset characterized by Severity Conditional Value-at-Risk (sCVaR_alpha). Contrasting internal activations across counterfactual rollout branches yields a sparse synapse attribution mask, over which we compute a closed-form weight update dW* right-projected onto the orthogonal complement (null space) of nominal activation covariance. Across a coupled three-scenario vehicle dynamics benchmark (6/64 = 9.4% weights edited, sCVaR_99%: 100.0 -> 0.0, worst-case SMaR_99%: -1.42 m -> +0.99 m, 0.01 m nominal drift) and an independent blind SGD-trained policy benchmark (290 parameters trained from scratch on dry asphalt, reducing held-out unseen tail crashes from 81.0% to 1.0% in 12.9 ms without the 2.94x nominal drift induced by dense SGD), CWSR eliminates coupled tail crashes with zero gradient descent epochs. Live Interactive Verification Benchmark: https://cwsr.unseenengine.tech/Research Platform: https://unseenengine.tech/Author Correspondence: [email protected] | [email protected]

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-12
DOI
https://doi.org/10.5281/zenodo.24178679
Primary Topic
Reinforcement Learning in Robotics
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Counterfactual Weight-Space Repair (CWSR): Mitigating Heavy-Tail Autonomous Driving Failures via Null-Space Projected Synapse Surgery

Andrew Turtsevych
Zenodo (CERN European Organization for Nuclear Research)
Reinforcement Learning in Robotics
preprint

Counterfactual Weight-Space Repair (CWSR): Mitigating Heavy-Tail Autonomous Driving Failures via Null-Space Projected Synapse Surgery

Andrew Turtsevych
preprint en

Abstract

Neural control policies trained via reinforcement learning or imitation learning achieve low error rates across nominal operating distributions (99.9% of routine trajectories), yet exhibit brittle failure modes in rare, high-severity tail regimes (0.1% edge cases such as abrupt tire-road friction drops on black ice mu = 0.14, dense-fog occluded pedestrian emergence, or asymmetric split-mu hydroplaning). Standard remedies—such as dense gradient fine-tuning (SGD/PPO) on synthetic counterexamples—modify 100% of network parameters globally, inducing catastrophic forgetting on nominal highway trajectories and creating conflicting control reflexes across coupled hazard scenarios. In this paper, we formulate Counterfactual Weight-Space Repair (CWSR), a non-gradient mechanistic intervention framework that localizes and patches a sparse subset (<11%) of causally responsible synapses directly in weight space. Using stochastic Monte Carlo trajectory ensembles under coupled nonlinear tire dynamics, we partition rollouts into a surviving safe subset characterized by positive Safety Margin at Risk (SMaR_alpha) and a heavy-tail failure subset characterized by Severity Conditional Value-at-Risk (sCVaR_alpha). Contrasting internal activations across counterfactual rollout branches yields a sparse synapse attribution mask, over which we compute a closed-form weight update dW* right-projected onto the orthogonal complement (null space) of nominal activation covariance. Across a coupled three-scenario vehicle dynamics benchmark (6/64 = 9.4% weights edited, sCVaR_99%: 100.0 -> 0.0, worst-case SMaR_99%: -1.42 m -> +0.99 m, 0.01 m nominal drift) and an independent blind SGD-trained policy benchmark (290 parameters trained from scratch on dry asphalt, reducing held-out unseen tail crashes from 81.0% to 1.0% in 12.9 ms without the 2.94x nominal drift induced by dense SGD), CWSR eliminates coupled tail crashes with zero gradient descent epochs. Live Interactive Verification Benchmark: https://cwsr.unseenengine.tech/Research Platform: https://unseenengine.tech/Author Correspondence: [email protected] | [email protected]

Zenodo (CERN European Organization for Nuclear Research)
Reinforcement Learning in Robotics
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon