Franka Panda: grasp-and-lift under simultaneous sensor noise + dynamics shift
Agilex MiniScout with uncertainty-guided adaptive model selection on edge
Most uncertainty-aware robotic systems collapse prediction uncertainty into a single scalar score and use it to trigger uniform corrective responses. This aggregation obscures whether uncertainty arises from corrupted observations or from mismatch between the learned model and the true system dynamics. As a result, corrective actions may be applied to the wrong component of the closed loop, degrading performance relative to leaving the policy unchanged.
We introduce a lightweight post hoc framework that decomposes uncertainty into aleatoric and epistemic components and uses these signals to regulate system responses at inference time. Aleatoric uncertainty is estimated from deviations in the observation distribution using a Mahalanobis density model, while epistemic uncertainty is detected using a noise-robust forward dynamics ensemble that isolates model mismatch from measurement corruption. The two signals remain empirically near-orthogonal during closed-loop execution and enable type-specific responses: high aleatoric uncertainty triggers observation recovery, while high epistemic uncertainty moderates control actions. The same signals also regulate adaptive perception by guiding model capacity selection during tracking inference.
Experiments demonstrate consistent improvements across both control and perception tasks. In robotic manipulation, the decomposed controller improves task success from 59.4% to 80.4% under compound perturbations and outperforms a combined uncertainty baseline by up to 21.0%. In adaptive tracking inference on MOT17, uncertainty-guided model selection reduces average compute by 58.2% relative to a fixed high-capacity detector while preserving detection quality within 0.4%.
Our framework operates as a post-hoc module that requires no retraining of the base policy. It decomposes uncertainty into two physically distinct components:
Estimated via Mahalanobis distance in observation space. Measures deviation from the nominal state distribution. When triggered, applies observation recovery by resampling the sensor (N=5 resamples, averaging).
Examples: encoder noise, calibration drift, occlusion, blur
Detected using a noise-robust forward dynamics ensemble (K=5 MLPs) trained with noise-augmented data paired with clean targets. When triggered, applies action dampening (scaling factor α=0.30).
Examples: mass change, friction variation, actuator drift, viewpoint shift
A key property is empirical orthogonality: the two signals maintain near-zero correlation (r=0.048 across 21,324 MOT17 detections), confirming they capture distinct disturbance mechanisms. This enables independent, type-specific interventions without cross-contamination.
Task success rate (%) on Isaac-Lift-Cube-Franka (N=1,000 episodes per condition). The Franka Emika Panda 7-DOF arm must reach, grasp, and lift a cube to a target height.
| Method | Nominal | Sensor Noise | Dynamics Shift | Compound |
|---|---|---|---|---|
| Vanilla (no intervention) | 100.0 | 63.8 | 72.4 | 44.6 |
| Observation Recovery Only | 97.4 | 89.2 | 64.8 | 57.3 |
| Action Dampen Only | 88.6 | 58.4 | 78.3 | 50.2 |
| Total-U (combined scalar) | 96.8 | 80.6 | 70.1 | 59.4 |
| Decomposed (Ours) | 99.6 | 94.2 | 84.2 | 80.4 |
| Δ vs. Total-U | +2.8 | +13.6 | +14.1 | +21.0 |
Total-U fails below vanilla under pure dynamics shift (70.1% vs. 72.4%) because it incorrectly applies observation recovery to clean sensor readings. Our decomposed approach avoids this cross-contamination.
(a) Aleatoric and epistemic trigger rates under four perturbation conditions. The aleatoric signal fires on 85.2% of steps under sensor perturbation while the epistemic signal fires on only 12.4% under dynamics shift — confirming near-independence. (b) Object height in a representative compound episode: Vanilla and Total-U fail to lift the cube above the 0.2 m success threshold, while Decomposed succeeds.
Uncertainty-guided adaptive model selection across YOLOv8 variants (Nano 3.2M to XLarge 68.2M params) on MOT17 (7 sequences, 4,746 frames).
| Sequence | Frames | Nano (%) | Small (%) | Medium (%) | Large (%) | XLarge (%) | Switches | Savings (%) |
|---|---|---|---|---|---|---|---|---|
| MOT17-02 | 550 | 18.2 | 7.3 | 35.3 | 32.0 | 7.2 | 12 | 58.3 |
| MOT17-04 | 1050 | 25.2 | 4.3 | 28.1 | 34.7 | 7.7 | 18 | 57.5 |
| MOT17-05 | 750 | 27.0 | 8.0 | 29.9 | 28.1 | 7.0 | 15 | 59.5 |
| MOT17-09 | 450 | 26.8 | 7.1 | 29.7 | 32.4 | 4.0 | 14 | 60.2 |
| MOT17-10 | 593 | 20.6 | 15.5 | 25.3 | 25.3 | 13.3 | 18 | 57.4 |
| MOT17-11 | 867 | 24.7 | 6.9 | 27.8 | 27.0 | 13.6 | 19 | 56.2 |
| MOT17-13 | 675 | 23.4 | 8.5 | 31.2 | 29.3 | 7.6 | 16 | 58.6 |
| Average | 4,746 | 23.7 | 8.2 | 29.6 | 29.8 | 8.6 | 16 | 58.2 |
The policy uses XLarge (the most expensive model) only 8.6% of the time, achieving 58.2% average compute savings while maintaining detection quality within 0.4% IoU of the fixed XLarge baseline.
Adaptive model selection for MOT17-04. Top: Temporal uncertainty evolution (epistemic in red, aleatoric in blue). Bottom: Selected YOLO models color-coded by capacity. The policy correlates model scaling with epistemic spikes while ignoring aleatoric elevation, validating orthogonality-aware learning.
Joint distribution across 21,324 MOT17 detections showing near-zero correlation (r=0.048), confirming the two uncertainties capture distinct mechanisms.
Per-sequence correlation heatmap on MOT17. Orthogonality between aleatoric and epistemic signals holds consistently across all sequences (r ≤ 0.05).
Our orthogonal decomposition achieves 58.2% average computational savings vs. 44.6% for total uncertainty (+13.6% improvement). Total uncertainty conservatively uses larger models when combined uncertainty is high, while our method recognizes that high aleatoric uncertainty alone does not require increased model capacity.
Franka Emika Panda performing lift-cube task under compound perturbations. The decomposed controller successfully completes the grasp-and-lift despite simultaneous sensor noise and dynamics shift.
Agilex MiniScout mobile robot with Intel NUC edge platform running adaptive uncertainty-guided model selection for real-time computer vision tasks.
@article{kumar2025uncertainty,
author = {Kumar, Divake and Tayebati, Sina and Naik, Devashri and Poggi, Patrick and Rios, Amanda Sofie and Ahuja, Nilesh and Trivedi, Amit Ranjan},
title = {TRIAGE: Type-Routed Interventions via Aleatoric-Epistemic Gated Estimation},
year = {2025},
eprint = {2603.08128},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2603.08128}
}