GUARD: Sensitivity-aware Uncertainty Attribution for Safe Traditional and Learning-based Robot Control
Authors: Woo-Jeong Baek, Johannes A. Gaus, Jonas Kiemel, Junheon Yoon, Suhan Park, Tamim Asfour, Jaeheung Park
Organizations: Hertie Institute for Clinical Brain Research & Center for Integrative Neuroscience, University of Tübingen, Germany. · Artificial Intelligence Institute (AIIS), Seoul National University, Republic of Korea. · School of Robotics, Kwangwoon University, Seoul, Republic of Korea. · Department of Intelligence and Information, Graduate School of Convergence Science and Technology, Seoul National University, Republic of Korea.
This paper presents the framework GUARD (Guided robot control via Uncertainty attribution and probAbilistic kernel optimization for Risk-aware Decision making) that combines a technique from probabilistic state estimation with an uncertainty identification method to derive critical thresholds for robot control algorithms. The first part focuses on complementing a method for probabilistic kernel optimization (PKO) by a sensitivity-aware uncertainty attribution technique. Next, the estimated uncertainties and their sources are forwarded to the robot controllers to adapt safety margins and trigger the initialization of safe backup trajectories. To validate GUARD's applicability and performance, experiments are conducted on a traditional model predictive control (MPC) algorithm and a reinforcement learning (RL) based controller. The results demonstrate that GUARD reduces collision rates by 44-67% in the RL-based evaluation, while decreasing backup policy activation by 50-75% in the medium- and high-uncertainty profiles. In the MPC simulation, GUARD successfully avoids collisions. By unifying the probabilistic kernel optimization algorithm from robot state estimation with the uncertainty estimation method for learning-based control, GUARD brings together traditional techniques and learning for safe and efficient robot applications.
Figures & tables
Fig. 1 : GUARD combines a kernel optimization technique with an uncertainty attribution method to develop safe robot control frameworks. The validation experiments on a MPC algorithm and a reinforcement learning (RL)-based controller show that GUARD successfully avoids collisions while preserving efficiency. Figures adapted from [ 1 ] , [ 2 ] , [ 3 ] .
Fig. 2 : Overview of State-of-the-Art literature and our contribution. While several collision avoidance methods have been presented in the robotic subfields, research on effectively combining methods from state estimation, control and perception remains sparse. GUARD provides one step toward the unification of methods by bringing together a probabilistic kernel optimization algorithm with an uncertainty attribution technique. The uncertainty with its source is fed into control frameworks to exploit the efficiency of learning while ensuring safety via explainable methods.
Fig. 3 : Schematic description of the PKO in [ 1 ] . PKO assigns weights to solve non-convex optimization problems.
Fig. 4 : Pipeline of the TCAV-GPC module: The TCAV-GPC algorithm provides uncertainty attributions via concept scores derived on the basis of the technique in [ 5 ] .
Fig. 5 : The RMPCC controller parametrizes the robot path via a path parameter s . The error is split into a contouring and a lag error by referring to the local path tangent. Fig. taken from [ 2 ] .
Fig. 6 : The pipeline of SafeMotionsRisk. The controller makes use of a task policy trained to complete a specific task and a backup policy trained to avoid collisions. If a defined threshold on the collision risk is exceeded, the controller switches to the backup policy to generate a collision-free trajectory. Adjusted from [ 3 ] .
Fig. 7 : Schematic description of GUARD. The point cloud is aligned via the PKO in [ 1 ] . The output of PKO combined with TCAV-GPC is used to extract relevant features and perform the sensitivity-aware uncertainty attribution method. Based on the estimated uncertainties, the collision margin of a traditional controller and the selected policy of a hybrid learning-based controller are adapted in real-time.
Fig. 8 : Comparison of residuals under injected uncertainty. Distribution of mean residuals for standard ICP, PKO-ICP, and Stein-ICP on synthetic shapes (HELIX, TORUS, KNOT). The PKO-ICP algorithm matches Stein-ICP accuracy trends while remaining faster and more accurate than standard ICP.
Method
Time/run
Helix
Torus
Knot
Std (CPU)
0.06–0.15 s
≈0.25
≈0.30
>0.80
PKO (CPU)
0.12–0.35 s
≈0.20
≈0.25
≈0.55
Stein (GPU)
0.80–1.00 s
≈0.15
≈0.20
≈0.30
TABLE I : Latency and mean residuals (approx.).
Fig. 9 : 7-DoF manipulator follows a 3D lemniscate with a dynamic obstacle; (a) setup; (b) minimum robot–obstacle distance for RMPCC (blue) and GUARD (red); dashed: safety margin εobs ; green boxes indicate collisions.
Profile
Abl.
Goal (%)
Collision (%)
Backup Frac.
Low
A0
96.0
9.5
0.033
A1
92.5
3.5
0.033
Medium
A0
49.0
6.0
0.067
A1
51.0
2.0
0.033
High
A0
12.0
10.5
0.133
A1
9.0
4.5
0.033
TABLE II : SafeMotionsRisk ablation results (200 deterministic episodes per condition). A0: Uncertainty value, A1: uncertainty value and source.
Department of Robotics, University of Michigan, Ann Arbor, MI, 48109 USA. · Department of Aerospace Engineering, University of Michigan, Ann Arbor, MI, 48109 USA.