Research map / Probabilistic and scientific modeling / Molecules and proteins
Conformal Prediction in The Loop: A Feedback-Based Uncertainty Model for Trajectory Optimization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Feedback-Based Conformal Prediction (Fb-CP) that feeds realized trajectories back into CP-based risk allocation for shrinking-horizon trajectory optimization under a joint chance constraint, with an iterative risk allocation (IRA) algorithm.
Problem
Existing CP-based trajectory optimization uses prediction regions sequentially, so realized decisions cannot feed back to adjust the regions, leaving conservatism in collision-avoidance planning under uncertain obstacle motion.
Approach
Compute a posterior violation probability from realized states using a held-out calibration subset, reclaim the unused risk (alpha minus beta) and reallocate it to future time steps, either evenly (ARA) or iteratively via an active/inactive constraint scheme (IRA), while preserving coverage guarantees.
Key results
- On the quadrotor model, Fb-CP-IRA reduces average cost by 58.50% relative to S-CP on average, and Fb-CP-ARA by 11.34%.
- On the kinematic vehicle model, Fb-CP-ARA reduces cost by 7.21% to 11.26% relative to S-CP across alpha in {0.05,...,0.20}.
- Fb-CP-IRA's collision avoidance rates are around 96-98% at alpha 0.05-0.15 and do not exceed the risk tolerance in the reported runs, though at alpha=0.20 the rate (91.2%) is still above the 80% implied level.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add feedback-based posterior risk adjustment using ARA | Sequential CP · Quadrotor model, α=0.05–0.20 | average cost: 17.321 -> 15.356 (α=0.05); 16.17 -> 14.228 (α=0.10); 14.83 -> 12.354 (α=0.15); 13.217 -> 10.22 (α=0.20) | Helped |
| use iterative rather than average risk allocation (ablation) | Fb-CP · Quadrotor model, α=0.05–0.20 | average cost: 15.356 -> 7.189 (α=0.05); 14.228 -> 6.798 (α=0.10); 12.354 -> 6.191 (α=0.15); 10.22 -> 5.398 (α=0.20) | Helped |
| add feedback-based posterior risk adjustment using ARA | Sequential CP · Kinematic vehicle model, α=0.05–0.20 | average cost: 22.20 -> 20.46 (α=0.05); 20.24 -> 18.78 (α=0.10); 19.22 -> 17.35 (α=0.15); 17.05 -> 15.13 (α=0.20) | Helped |
| use iterative rather than average risk allocation (ablation) | Fb-CP · Kinematic vehicle model, α=0.05–0.20 | average cost: 20.46 -> 4.77 (α=0.05); 18.78 -> 3.52 (α=0.10); 17.35 -> 3.18 (α=0.15); 15.13 -> 2.89 (α=0.20) | Helped |
| use iterative rather than average risk allocation (ablation) | Fb-CP · Quadrotor model | average computation time: 0.022 -> 0.038; 0.020 -> 0.039; 0.021 -> 0.037; 0.020 -> 0.036 | Hurt |
| add feedback-based posterior risk adjustment using ARA | Sequential CP · Kinematic vehicle model | average computation time: 0.111 -> 0.100 (α=0.05); 0.093 -> 0.087 (α=0.10); 0.078 -> 0.085 (α=0.15); 0.076 -> 0.078 (α=0.20) | No clear effect |
Surprising findings
- Fb-CP-IRA is substantially more costly in computation time (about 0.13 s vs 0.11 s for S-CP on vehicle model) but yields much lower cost, suggesting decision-focused reallocation drives gains.
Limitations
- IRA requires iterative solves of the trajectory optimization, increasing computation time.
- Calibration data is split into two subsets, reducing effective sample size for each part.
- Posterior risk guarantee is probabilistic and only asymptotically equals the prior bound.
Experimental setup
Datasets: TrajNet++ generated obstacle trajectories Stanford Drone Dataset
Benchmarks: kinematic vehicle model simulation linear quadrotor simulation dynamic bicycle model Stanford Drone Dataset
Metrics: average cost computation time collision avoidance rate
Baselines: Conformal Control (CC) ACI for Motion Planning (ACI-MP) Recursively Feasible MPC using CP (RF-CP) Sequential CP (S-CP) Fb-CP with ARA
Builds on: Conformal prediction Sequential CP Iterative risk allocation ACI
Closest papers by meaning
- Learning Polyhedral Conformal Sets for Robust Optimization (NeurIPS 2026 · similarity 0.87)
- Optimal transport-based conformal prediction (ICML 2025 · similarity 0.86)
- Conformalized Adaptive Forecasting of Heterogeneous Trajectories (ICML 2024 · similarity 0.86)
- CUQDS: Conformal Uncertainty Quantification Under Distribution Shift for Trajectory Prediction (AAAI 2025 · similarity 0.85)
- AdaptNC: Adaptive Nonconformity Scores for Conformal Prediction under Distribution Shift (NeurIPS 2026 · similarity 0.85)
- Multi-model Ensemble Conformal Prediction in Dynamic Environments (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.