Learnable Conformal Prediction for Safe and Efficient Robotics under Perception and Planning Uncertainties
Divake Kumar, Sina Tayebati, Francesco Migliarba, Ranganath Krishnan, Amit Ranjan Trivedi
AI summary
Problem
Standard conformal prediction relies on fixed nonconformity scores that ignore contextual cues, often producing overly conservative or unsafe uncertainty intervals for robotics. This limits real-world deployment where risk depends heavily on situational context and computational constraints.
Approach
LCP trains a lightweight neural network to generate adaptive nonconformity scores from geometric, semantic, and model-derived features, dynamically balancing coverage and efficiency. This learned scoring function is calibrated on held-out data to restore finite-sample statistical guarantees without requiring ensembles or repeated inference.
Key results
- 91.5% navigation success rate with only 4.5% path inflation on MRPB
- 46–54% reduction in object detection interval width at 90% coverage
- 4.7–9.9% shrinkage of classification prediction sets across multiple benchmarks
- Real-time deployment on resource-constrained edge hardware with minimal overhead
Why it matters
Provides a statistically rigorous, context-aware uncertainty quantification method that enables safer and more efficient robotic systems on real-world edge hardware.
Abstract
Deep learning models in robotics often output point estimates with poorly calibrated confidences, offering no native mechanism to quantify predictive reliability under novel, noisy, or out-of-distribution inputs. Conformal prediction (CP) addresses this gap by providing distribution-free coverage guar- antees, yet its reliance on fixed nonconformity scores ignores context and can yield intervals that are overly conservative or unsafe. We address this with Learnable Conformal Prediction (LCP), which replaces fixed scores with a lightweight neural function sθ(x) = fθ(φ(x)) that leverages geometric, semantic, and model cues. Trained to balance coverage, efficiency, and calibration, LCP preserves CP’s finite-sample guarantees while producing intervals that adapt to instance difficulty, achieving context-aware uncertainty without ensembles or repeated infer- ence. On the MRPB benchmark, LCP raises navigation success to 91.5% versus 87.8% for Standard CP, while limiting path inflation to 4.5% compared to 12.2%. For object detection on COCO, BDD100K, and Cityscapes, it reduces mean interval width by 46–54% at 90% coverage, and on classification tasks (CIFAR-100, HAM10000, ImageNet) it shrinks prediction sets by 4.7–9.9%. The method is also computationally efficient, achieving real-time performance on resource-constrained edge hardware (Intel NUC with footprint 4.6 × 4.4 inch2 and idle power < 30 W) while simultaneously providing uncertainty estimates along with the mean prediction. Project page: [Code, Video & Results]