Adaptive self-supervised learning for real-time problem solving in autonomous systems.
TL;DR
The AdaptSolveAI with SelfSolveNet framework provides a practical recipe for resilient and safety-aware autonomy by enabling continuous learning from unlabeled streams, adaptation within fixed computational budgets, and action execution under control barrier function-based safety constraints.
Abstract
Autonomous systems must perform safely and efficiently when deployed in the real world, where data distributions drift and no labels are available during deployment. Current approaches either assume stationarity, adapt slowly (and only through sufficiently narrow objectives), or leave learning and explicit safety uncoupled (neglecting stringent latency budgets). We present the safety-aligned, latency-bounded problem solving and continuous, label-free adaptation framework, AdaptSolveAI with SelfSolveNet. It combines a multi-objective self-supervised paradigm (contrastive, masked reconstruction, cross-modal alignment, distillation, and a head informed by the barrier), a shift-aware prioritized replay (novelty, safety margin, Fisher sensitivity), and parameter-efficient updates (adaptors/normalization/low-rank layers) under the hard constraint of a compute budget. That is, a control-barrier safety supervisor minimally corrects actions, and is given the same safety-signal used in the representation learning stage. On driving, navigation, manipulation and scheduling benchmarks with S1-S3 distribution shifts, AdaptSolveAI outperforms the strongest baseline on normalized return (0.86 ± 0.01 vs. 0.79 ± 0.02), success rate (77.9%±1.2 vs. 70.3%±1.3), and safety violations (3.9%±0.3 vs. 5.8%±0.3), while maintaining FPS (frames per second) and deadline missed (%) (≤ 1% of deadlines missed). Time-to-adapt is less (310 ± 20 vs. 520 ± 30 steps) The area under the robustness curve is also higher (0.857 vs. 0.776) under progressive camera dropout, with recovery after sensor restoration taking place faster. Abalations show the performance benefits cannot be achieved without the safety-aware and cross-modal heads, prioritized replay, and adapters. The framework provides a practical recipe for resilient and safety-aware autonomy by enabling continuous learning from unlabeled streams, adaptation within fixed computational budgets, and action execution under control barrier function-based safety constraints. It is directly applicable to embedded deployments and offers a path toward standardized evaluation and certification of learning-enabled systems.