GUARD: Sensitivity-aware Uncertainty Attribution for Safe Traditional and Learning-based Robot Control
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: We're looking at the paper: "GUARD: Sensitivity-aware Uncertainty Attribution for Safe Traditional and Learning-based Robot Control".
Dev: Here's the gist of the paper. This paper presents GUARD (Guided robot control via Uncertainty attribution and probabilistic kernel optimization for Risk-aware Decision making) that combines traditional control with an uncertainty-aware perception technique using active learning with real-time capability for safe robot collision avoidance.
Rosa: By doing so, this manuscript addresses the central challenge in robotics of finding a reasonable compromise between traditional methods and learning algorithms to foster the development of safe, yet efficient and flexible applications.
Dev: The core idea is that safety is evaluated inversely by the uncertainty.
Rosa: Framework Components The GUARD framework unifies reactive model predictive contouring control (RMPCC) with an Iterative Closest Point (ICP) algorithm that enables the attribution of uncertainty sources online using active learning with real-time capability via a probabilistic kernel optimization technique.
Title and authors: Rosa: So we're looking at this paper called "GUARD: Sensitivity-aware Uncertainty Attribution for Safe Traditional and Learning-based Robot Control" by Johannes Gaus, Junheon Yoon, Woo-Jeong Baek, Seungwon Choi, Suhan Park, and Jaeheung Park. It sounds a bit dense with all those technical terms.
Dev: Yeah it is. The title itself points to combining traditional control with uncertainty attribution for safe robot systems. It suggests they are tackling that big problem of finding a balance between old methods and new learning algorithms in robotics, which is what the paper aims for >
Rosa: Exactly. Basically, they're suggesting that safety in these robot applications should be evaluated inversely by the uncertainty itself. Instead of assuming perfect perception or perfect control, you use the uncertainty to guide how safe things need to be >
Taro: That makes sense from a real-world standpoint. If the robot is really unsure about where an object is, it should probably act more cautiously, right? It's about handling when things go wrong in an unpredictable way >
Dev: Right. And this paper proposes a framework called GUARD to unify reactive model predictive contouring control with an uncertainty aware Iterative Closest Point algorithm that uses active learning for real-time capability >
The paper's summary: Rosa: So, the core idea of this GUARD framework is to use uncertainty attribution and probabilistic kernel optimization for risk-aware decision making to achieve safe robot collision avoidance >
Dev: It unifies reactive model predictive contouring control with an Iterative Closest Point algorithm that can attribute uncertainty sources online using active learning with real-time capability through a probabilistic kernel optimization technique >
Rosa: It's about taking those uncertainties from perception and feeding them directly into the reactive model predictive contouring control to adapt the safety conditions of the control barrier functions >
Taro: So, if perception is shaky, the control system automatically adjusts its safety margins online based on how uncertain it is about things >
Dev: Right. They have a pipeline that goes from environment perception to PKO Iterative Closest Point feature extraction and then kernel optimization and explainability before finally improving the control >
Rosa: It sounds like they're trying to make the system smarter about when it needs to be conservative versus when it can be efficient >
The paper's improvements: Dev: One key improvement mentioned is the active PKO via Concept Feedback mechanism which dynamically updates the tuner to a vector theta k that might include something like a correspondence gate or a robust loss scale >
Rosa: They only trigger this adaptation if necessary, specifically when they see concepts indicating why the estimate is uncertain, like when the dominant concept and or non-trivial uncertainty occur at most one tau on and the overall uncertainty is above some minimum threshold >
Taro: So it's not constantly changing things; it only adjusts when there's a specific reason to believe something is fuzzy, which sounds much more practical for real-time systems >
Dev: And they project these updated parameters onto the feasible set, which means whatever they change still has to stay within the limits of what the robot can safely do >
Rosa: Then there's this risk-aware control integration where the uncertainty modulates things like the contouring control barrier function relaxation, collision margins and desired speed >
Taro: So if that uncertainty spikes, it means they're going to automatically trade efficiency for conservatism by increasing slack penalties at high risk levels or reducing the desired speed >
Conclusion: Rosa: To wrap up this discussion on GUARD, the paper successfully combines traditional control with learning-driven perception to generate safe robot movements through uncertainty-aware ICP and risk-adaptive RMPCC >
Dev: It shows that by optimizing kernel parameters alongside uncertainty aware perception, they can derive a bounded risk measure rho k from concept posteriors and overall uncertainty using the formula clip zero one beta zero u k plus sigma c in C beta c pi k of c >
Taro: That risk measure essentially linearly inflates the margins and increases the slack penalty while reducing speed to tighten the safe set when uncertainty is high >
Rosa: And they flag out-of-distribution frames using a reject rule where if the overall uncertainty is above tau u or if the maximum concept posterior probability falls below tau p i, it triggers freezing kernel updates and reducing desired speed while inflating margins >
Dev: So, GUARD unifies reactive model predictive contouring control with uncertainty aware ICP to generate safe robot movements by adapting controller margins while preserving real-time performance >
Taro: It proves that this approach can adapt the controller settings to ensure safety without losing the real-time performance they need for sudden disturbances or obstacles >
Hertie Institute for Clinical Brain Research & Center for Integrative Neuroscience, University of Tubingen, Germany · Department of Intelligence and Information, Graduate School of Convergence Science and Technology, Seoul National University, Republic of Korea · Artificial Intelligence Institute (AIIS), Seoul National University, Republic of Korea · Department of Naval Architecture and Ocean Engineering, Seoul National University, Republic of Korea · School of Robotics, Kwangwoon University
cs.RO
Submitted: 2025-09-27
Updated: 2026-11-05
Comments: 9 pages, 9 figures, 2 tables. Substantially revised and expanded version with updated title and authors. Extended validation using model predictive control and reinforcement learning-based control
License: http://creativecommons.org/licenses/by-nc-sa/4.0/
Importance score: 87/100
The gist: This paper presents GUARD (Guided robot control via Uncertainty attribution and probabilistic kernel optimization for Risk-aware Decision making) that combines traditional control with an…
Key concepts
- RMPCC
- Reactive Model Predictive Contouring Control is a method that follows a desired path while ensuring safety constraints are met. It uses control barrier functions (CBFs) to guarantee collision avoidance and smooth evasive movements, operating in real-time at 100Hz.
- ICP
- Iterative Closest Point (ICP) is an algorithm used for point cloud registration, which aligns two sets of 3D points. GUARD uses ICP to estimate robot pose and simultaneously extracts uncertainty information about the alignment quality.
- Probabilistic Kernel Optimization (PKO)
- PKO adapts the robust kernel parameters in ICP by matching the observed data residuals with a model-based distribution. This makes ICP more robust against noise and partial overlaps, improving registration accuracy through iterative refinement.
- Uncertainty Attribution
- This process identifies *why* an estimate is uncertain by classifying it into specific concepts (like sensor noise or pose error) using a Gaussian Process Classifier. This allows the system to understand the source of uncertainty for better risk management.
Terminology
Summary
This paper presents GUARD (Guided robot control via Uncertainty attribution and probabilistic kernel optimization for Risk-aware Decision making) that combines traditional control with an uncertainty-aware perception technique using active learning with real-time capability for safe robot collision avoidance. p. 0 GUARD: Toward a Compromise between Traditional Control and
Learning for Safe Robot Systems
Johannes A. Gaus1
, Junheon Yoon2
, Woo-Jeong Baek3
, Seungwon Choi4
, Suhan Park5
, Jaeheung Park2
Abstract— This paper presents the framework GUARD
(Guided robot control via Uncertainty attribution and
probAbilistic kernel optimization for Risk-aware Decision making) that combines traditional control with an uncertaintyaware perception technique using active learning with realtime capability for safe robot collision avoidance. Gaus1
, Junheon Yoon2
, Woo-Jeong Baek3
, Seungwon Choi4
, Suhan Park5
, Jaeheung Park2
Abstract— This paper presents the framework GUARD
(Guided robot control via Uncertainty attribution and
probAbilistic kernel optimization for Risk-aware Decision making) that combines traditional control with an uncertaintyaware perception technique using active learning with realtime capability for safe robot collision avoidance. By doing so,
this manuscript addresses the central challenge in robotics of
finding a reasonable compromise between traditional methods
and learning algorithms to foster the development of safe, yet
efficient and flexible applications. Gaus1
, Junheon Yoon2
, Woo-Jeong Baek3
, Seungwon Choi4
, Suhan Park5
, Jaeheung Park2
Abstract— This paper presents the framework GUARD
(Guided robot control via Uncertainty attribution and
probAbilistic kernel optimization for Risk-aware Decision making) that combines traditional control with an uncertaintyaware perception technique using active learning with realtime capability for safe robot collision avoidance. By doing so,
this manuscript addresses the central challenge in robotics of
finding a reasonable compromise between traditional methods
and learning algorithms to foster the development of safe, yet
efficient and flexible applications. For example, it remains unclear how the
perception parameters affect the robot control performance in
the context of safety. To complement traditional techniques
with learning in a safe and efficient manner, this paper
contends that safety is evaluated inversely by the uncertainty.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Framework Components
The GUARD framework unifies reactive model predictive contouring control (RMPCC) with an Iterative Closest Point (ICP) algorithm that enables the attribution of uncertainty sources online using active learning with real-time capability via a probabilistic kernel optimization technique. This unification adapts the safety conditions of the CBFs in the RMPCC by referring to the perception uncertainties. p. 0 By further
actively adapting the kernel parameters via our previous work
in [5], robustness is achieved. This unification adapts the
safety conditions of the CBFs in the RMPCC by referring to
the perception uncertainties. By further
actively adapting the kernel parameters via our previous work
in [5], robustness is achieved. This unification adapts the
safety conditions of the CBFs in the RMPCC by referring to
the perception uncertainties. THE FRAMEWORK GUARD
This section presents GUARD (Guided robot control via
Uncertainty attribution and probAbilistic kernel optimization
for Risk-aware Decision making) by combining and
extending methods from prior works for safe robot systems. The following steps, illustrated in the figure 3, outline the
complete GUARD pipeline:
- Env. The computation O(|C|d) is negligible
compared to ICP and feature extraction. Uncertainty-aware Robot Control
This section integrates the perception module into the RMPCC to relax the assumption on the perfectly accurate
perception in [4].Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Active Adaptation Mechanism
The active PKO via Concept Feedback mechanism dynamically updates the tuner to a vector θk that may include a correspondence gate or robust-loss scale. The adaptation is triggered only if necessary via the concepts that indicate why the estimate is uncertain, specifically when GUARD
selects the exact schedules empirically. The updated parameters are projected onto the feasible set Θ:
θk+1 = ΠΘ
θk + ∆θk
.dominant concept and/or non-trivial uncertainty occur (maxc πk(c) ≥ τon and uk ≥ umin)
. The updated parameters are projected onto the feasible set Θ. p. 2Where the paper says it
Risk-Aware Control Integration
The uncertainty-aware robot control integrates perception into RMPCC to relax the assumption on perfectly accurate perception. The scalar risk and per-concept posteriors modulate (i)the CBF relaxation (large slack penalties at high risk), (ii)collision margins and desired speed vdes, and (iii)soft weights wc,wl,wo,wvs to trade efficiency for conservatism if perception uncertainty spikes. p. 2 Uncertainty-aware Robot Control
This section integrates the perception module into the RMPCC to relax the assumption on the perfectly accurate
perception in [4]. Since uncertainties are proportional to
risks, the perception uncertainty enables to develop riskaware robot controllers: The scalar risk and per-concept posteriors modulate (i)the CBF relaxation (large slack penalties
at high risk), (ii)collision margins and desired speed vdes,
and (iii)soft weights wc,wl
,wo,wvs
to trade efficiency for
conservatism if perception uncertainty spikes. The controller becomes conservative for high uncertainties. Calibrated concept posteriors πk(c) and entropy uk yield
ρk = clip[0,1]
β0 uk +
X
c∈C
βc πk(c)
, (8)
where β ≥0 weight each source.Where the paper says it
Where the paper says it
Experimental Validation
Experiments evaluate standard point-to-plane ICP, Stein-ICP, and PKO-ICP on synthetic benchmarks like HELIX, TORUS, and KNOT. PKO-ICP shows a promising accuracy–latency trade-off compared to standard ICP. p. 2 PKO-ICP attains a promising accuracy–latency
trade-off, see Tab. I: Standard ICP runs at ∼0. OOD Robustness. GUARD flags out-of-distribution frames
using a reject rule on the calibrated uncertainty:
OODk = 1
h
uk > τu ∨ max
c∈C
πk(c) < τp
i
.Where the paper says it
Where the paper says it
Conclusion
GUARD unifies traditional control RMPCC with learning-driven perception to generate safe robot movements. The framework successfully combines the optimization of kernel parameters with uncertainty-aware perception to generate safe robot movements. p. 3 CONCLUSION
This paper presented GUARD that unifies traditional control
RMPCC in [4] with learning-driven perception. GUARD
combines the optimization of kernel parameters with
uncertainty-aware perception to generate safe robot movements. The attributed risk adapts controller margins, yielding
safety while preserving real-time performance. Experimental results demonstrate the performance and reliability of
GUARD. GUARD
combines the optimization of kernel parameters with
uncertainty-aware perception to generate safe robot movements. The attributed risk adapts controller margins, yielding
safety while preserving real-time performance.Where the paper says it
Where the paper says it
Where the paper says it
--- Page 1 ---
GUARD: Toward a Compromise between Traditional Control and Learning for Safe Robot Systems Johannes A. Gaus1, Junheon Yoon2, Woo-Jeong Baek3, Seungwon Choi4, Suhan Park5, Jaeheung Park2 Abstract— This paper presents the framework GUARD (Guided robot control via Uncertainty attribution and probAbilistic kernel optimization for Risk-aware Decision making) that combines traditional control with an uncertaintyaware perception technique using active learning with realtime capability for safe robot collision avoidance. p. 0 GUARD: Toward a Compromise between Traditional Control and
Learning for Safe Robot Systems
Johannes A. Gaus1
, Junheon Yoon2
, Woo-Jeong Baek3
, Seungwon Choi4
, Suhan Park5
, Jaeheung Park2
Abstract— This paper presents the framework GUARD
(Guided robot control via Uncertainty attribution and
probAbilistic kernel optimization for Risk-aware Decision making) that combines traditional control with an uncertaintyaware perception technique using active learning with realtime capability for safe robot collision avoidance. By doing so,
this manuscript addresses the central challenge in robotics of
finding a reasonable compromise between traditional methods
and learning algorithms to foster the development of safe, yet
efficient and flexible applications. By unifying a reactive model
predictive countouring control (RMPCC) with an Iterative
Closest Point (ICP) algorithm that enables the attribution of
uncertainty sources online using active learning with real-time
capability via a probabilistic kernel optimization technique,
GUARD inherently handles the existing ambiguity of the term
safety that exists in robotics literature. By unifying a reactive model
predictive countouring control (RMPCC) with an Iterative
Closest Point (ICP) algorithm that enables the attribution of
uncertainty sources online using active learning with real-time
capability via a probabilistic kernel optimization technique,
GUARD inherently handles the existing ambiguity of the term
safety that exists in robotics literature. Experimental studies
indicate the high performance of GUARD, thereby highlighting
the relevance and need to broaden its applicability in future. Experimental studies
indicate the high performance of GUARD, thereby highlighting
the relevance and need to broaden its applicability in future. INTRODUCTION
Developing safe and flexible robot applications is a central,
yet open issue in robotics. INTRODUCTION
Developing safe and flexible robot applications is a central,
yet open issue in robotics. One major difficulty arises due
to the challenge to develop systems that are sufficiently
explainable to verify safety, and enable the use of efficient
learning methods. One major difficulty arises due
to the challenge to develop systems that are sufficiently
explainable to verify safety, and enable the use of efficient
learning methods. Thus, a suitable compromise must be
found such that safety, that requires explainability, can be
evaluated while preserving the benefits of learning. Thus, a suitable compromise must be
found such that safety, that requires explainability, can be
evaluated while preserving the benefits of learning. While
end-to-end learning like in ( [1], [2], [3]) have led to
promising results regarding their efficiency, their black-box
character causes a burden on assessing safety. While
end-to-end learning like in ( [1], [2], [3]) have led to
promising results regarding their efficiency, their black-box
character causes a burden on assessing safety. A second
challenge lies in the ambiguity of the safety definitions in the
different subdomains. A second
challenge lies in the ambiguity of the safety definitions in the
different subdomains. Robotic systems employ methods from
control, perception, and/or state estimation. Robotic systems employ methods from
control, perception, and/or state estimation. Especially, the
respective assumptions that must be met (e. Especially, the
respective assumptions that must be met (e. , control barrier
functions (CBFs) in traditional control, registration accuracy
in perception,. , control barrier
functions (CBFs) in traditional control, registration accuracy
in perception,. ) are formulated in a domain-specific manner. ) are formulated in a domain-specific manner. However, the relationship between these requirements is
often not known. However, the relationship between these requirements is
often not known. For example, it remains unclear how the
perception parameters affect the robot control performance in
the context of safety. For example, it remains unclear how the
perception parameters affect the robot control performance in
the context of safety. To complement traditional techniques
with learning in a safe and efficient manner, this paper
contends that safety is evaluated inversely by the uncertainty. To complement traditional techniques
with learning in a safe and efficient manner, this paper
contends that safety is evaluated inversely by the uncertainty. 1 Hertie Institute for Clinical Brain Research & Center for Integrative
Neuroscience, University of Tubingen, Germany. 1 Hertie Institute for Clinical Brain Research & Center for Integrative
Neuroscience, University of Tubingen, Germany. ¨
2 Department of Intelligence and Information, Graduate School of
Convergence Science and Technology, Seoul National University, Republic
of Korea. ¨
2 Department of Intelligence and Information, Graduate School of
Convergence Science and Technology, Seoul National University, Republic
of Korea. 3 Artificial Intelligence Institute (AIIS), Seoul National University, Republic of Korea. 3 Artificial Intelligence Institute (AIIS), Seoul National University, Republic of Korea. 4 Department of Naval Architecture and Ocean Engineering, Seoul
National University, Republic of Korea. 4 Department of Naval Architecture and Ocean Engineering, Seoul
National University, Republic of Korea. 5 School of Robotics, Kwangwoon University, Seoul, Republic of Korea. 5 School of Robotics, Kwangwoon University, Seoul, Republic of Korea. Thus, the uncertainty is treated as the key measure for safety
evaluation. Thus, the uncertainty is treated as the key measure for safety
evaluation. The uncertainty reflects the possible deviation of
a quantity and is expressed as a probability. The uncertainty reflects the possible deviation of
a quantity and is expressed as a probability. Logically, the
absence of uncertainties (uncertainty = 0) ensures safety: In
case all system variables are known with perfect accuracy at
each time step, dangerous situations can be avoided by introducing risk-avoiding measures in time. Logically, the
absence of uncertainties (uncertainty = 0) ensures safety: In
case all system variables are known with perfect accuracy at
each time step, dangerous situations can be avoided by introducing risk-avoiding measures in time. This paper suggests
GUARD to combine reactive model predictive contouring
control (RMPCC) from prior work in [4] that ensures smooth
evasive movements with uncertainty-aware ICP. This paper suggests
GUARD to combine reactive model predictive contouring
control (RMPCC) from prior work in [4] that ensures smooth
evasive movements with uncertainty-aware ICP. By further
actively adapting the kernel parameters via our previous work
in [5], robustness is achieved. This unification adapts the
safety conditions of the CBFs in the RMPCC by referring to
the perception uncertainties. Practically, more conservative
robot movements are generated in case of high perception
uncertainties while maximizing the efficiency for negligible
uncertainties. Practically, more conservative
robot movements are generated in case of high perception
uncertainties while maximizing the efficiency for negligible
uncertainties. The real-time capability (100Hz) enables to
handle sudden occurrence of obstacles or unexpected disturbances. The real-time capability (100Hz) enables to
handle sudden occurrence of obstacles or unexpected disturbances. a) Related Work: Classical path following can
be split into time–parameterized planners/trackers
and path–parameterized MPC. a) Related Work: Classical path following can
be split into time–parameterized planners/trackers
and path–parameterized MPC. The former optimize a
time–indexed trajectory that a low–level controller tracks
[6]–[11]; the latter optimizes progress along a spatial path to
gain temporal flexibility and real–time reactivity evaluated on
cars, UAVs, and manipulators [12], [13]. The former optimize a
time–indexed trajectory that a low–level controller tracks
[6]–[11]; the latter optimizes progress along a spatial path to
gain temporal flexibility and real–time reactivity evaluated on
cars, UAVs, and manipulators [12], [13]. For safety, artificial
potential fields have been widely used, while CBFs yield
forward–invariance guarantees and typically less intrusive
avoidance [14], [15]. For safety, artificial
potential fields have been widely used, while CBFs yield
forward–invariance guarantees and typically less intrusive
avoidance [14], [15]. RMPCC combines path–parameterized
MPC with CBFs and achieves high servo rates via
Jacobian linearization and Gauss–Newton/SQP [4]. RMPCC combines path–parameterized
MPC with CBFs and achieves high servo rates via
Jacobian linearization and Gauss–Newton/SQP [4]. On the
perception side, ICP remains the workhorse for point–cloud
registration [16], [17]. On the
perception side, ICP remains the workhorse for point–cloud
registration [16], [17]. Uncertainty has been analyzed
via closed-form covariances [18] and their limitations
[19], learned covariance predictors (CELLO-3D) [20], and
probabilistic/multi-modal variants (e. Uncertainty has been analyzed
via closed-form covariances [18] and their limitations
[19], learned covariance predictors (CELLO-3D) [20], and
probabilistic/multi-modal variants (e. , Stein-ICP) [21]. , Stein-ICP) [21]. However, most pipelines quantify how much uncertainty
exists but not why. However, most pipelines quantify how much uncertainty
exists but not why. Concept-based explanations as TCAV
[22] with point-cloud encoders (PointNet/DGCNN) [23],
[24] enable human-interpretable attribution. Concept-based explanations as TCAV
[22] with point-cloud encoders (PointNet/DGCNN) [23],
[24] enable human-interpretable attribution. Recent post-hoc
attribution for ICP exists but is computationally heavy
and not integrated with control [25]. Recent post-hoc
attribution for ICP exists but is computationally heavy
and not integrated with control [25]. Probabilistic Kernel
Optimization (PKO) adapts robust kernels by matching
model/data inlier distributions to improve registration
robustness [5]. Probabilistic Kernel
Optimization (PKO) adapts robust kernels by matching
model/data inlier distributions to improve registration
robustness [5]. Contribution. Contribution. GUARD differs by (i) attaching per-concept,
calibrated attributions to ICP uncertainty and (ii) feeding
these uncertainties into RMPCC to adapt safety margins
arXiv:2509. p. 0 Contribution. GUARD differs by (i) attaching per-concept,
calibrated attributions to ICP uncertainty and (ii) feeding
these uncertainties into RMPCC to adapt safety margins
arXiv:2509. GUARD differs by (i) attaching per-concept,
calibrated attributions to ICP uncertainty and (ii) feeding
these uncertainties into RMPCC to adapt safety margins
arXiv:2509. 23312v1 [cs.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 2 ---
online while preserving real-time performance, bridging perception failures to control guarantees. II. PRELIMINARIES AND PRIOR WORK Iterative Closest Point (ICP) Algorithm ICP estimates the rigid motion T = (R, t) ∈ SE(3) that aligns a source point cloud to a target cloud by alternating correspondence assignment and a small weighted least-squares update. p. 1 online while preserving real-time performance, bridging perception failures to control guarantees. PRELIMINARIES AND PRIOR WORK
Iterative Closest Point (ICP) Algorithm ICP estimates
the rigid motion T = (R, t) ∈ SE(3) that aligns a source
point cloud to a target cloud by alternating correspondence
assignment and a small weighted least-squares update. PRELIMINARIES AND PRIOR WORK
Iterative Closest Point (ICP) Algorithm ICP estimates
the rigid motion T = (R, t) ∈ SE(3) that aligns a source
point cloud to a target cloud by alternating correspondence
assignment and a small weighted least-squares update. In
the point-to-plane variant [16], each iteration minimizes
residuals of the form
ri = n
⊤
i
Rpi + t − qi
. In
the point-to-plane variant [16], each iteration minimizes
residuals of the form
ri = n
⊤
i
Rpi + t − qi
. The per-iteration residual set {ri}, inlier counts, and overlap
proxies provide both the pose estimate and rich cues about
the scene (e. The per-iteration residual set {ri}, inlier counts, and overlap
proxies provide both the pose estimate and rich cues about
the scene (e. , occlusion, poor overlap, dynamic objects). , occlusion, poor overlap, dynamic objects). Probabilistic Kernel Optimization (PKO)
Our previous work [5] uses above residuals to adapt the
robust kernel parameters by matching an inlier distribution
with the empirical residual. Probabilistic Kernel Optimization (PKO)
Our previous work [5] uses above residuals to adapt the
robust kernel parameters by matching an inlier distribution
with the empirical residual. To be specific, PKO refers to
the iteratively reweighted least squares (IRLS) algorithm
that employs a weighting scheme for the residuals by assigning smaller weights to outliers, thereby minimizing their
influence on the optimization. To be specific, PKO refers to
the iteratively reweighted least squares (IRLS) algorithm
that employs a weighting scheme for the residuals by assigning smaller weights to outliers, thereby minimizing their
influence on the optimization. PKO enhances the robustness
as follows: For the IRLS weight w(r; c) with the scaling
parameter c of the robust kernel, a model-based inlier probability Pmodel(inlier | r, c) = w(r; c)/Zw, and a data-based
inlier probability Pdata(inlier | r) from a short window of
recent residual magnitudes is defined. PKO enhances the robustness
as follows: For the IRLS weight w(r; c) with the scaling
parameter c of the robust kernel, a model-based inlier probability Pmodel(inlier | r, c) = w(r; c)/Zw, and a data-based
inlier probability Pdata(inlier | r) from a short window of
recent residual magnitudes is defined. At each IRLS iteration,
c is adapted via the Jensen-Shannon divergence JS:
c
⋆ = argmin
c
JS
Pdata(·) ∥ Pmodel(· | c)
. At each IRLS iteration,
c is adapted via the Jensen-Shannon divergence JS:
c
⋆ = argmin
c
JS
Pdata(·) ∥ Pmodel(· | c)
. This aligns the effective inlier boundary with the sensor data
reflecting the actual state. This aligns the effective inlier boundary with the sensor data
reflecting the actual state. With the improving alignment, c
⋆
decreases, reproducing a graduated-nonconvexity behavior. With the improving alignment, c
⋆
decreases, reproducing a graduated-nonconvexity behavior. PKO improves the tolerance against initialization uncertainties, sensor noise and partial overlaps. PKO improves the tolerance against initialization uncertainties, sensor noise and partial overlaps. TCAV–GPC for uncertainty attribution for ICP
The lack of interpretability in residual-driven adaptation
is addressed by integrating the technique of Testing with
Concept Activation Vectors (TCAV) [22]. TCAV–GPC for uncertainty attribution for ICP
The lack of interpretability in residual-driven adaptation
is addressed by integrating the technique of Testing with
Concept Activation Vectors (TCAV) [22]. The perception is
performed via RGB-D frames that are processed with ICP. The perception is
performed via RGB-D frames that are processed with ICP. With the obtained point clouds, TCAV extracts expressive
features via a pretrained model like Dynamic Graph CNN
(DGCNN) or a PointNet approach [23], [24]. With the obtained point clouds, TCAV extracts expressive
features via a pretrained model like Dynamic Graph CNN
(DGCNN) or a PointNet approach [23], [24]. For the embedding ek = fθ(Pk) generated by the encoder corresponding to
the point cloud Pk, a feature vector
ϕk =
ek ] ∈ R
d
is formed. For the embedding ek = fθ(Pk) generated by the encoder corresponding to
the point cloud Pk, a feature vector
ϕk =
ek ] ∈ R
d
is formed. By training a multi-class Gaussian Process Classifier (GPC) with these vectors on artificially perturbed data,
the goal is to classify ϕk
in three semantic uncertainty
concepts
C = {sensor noise, pose error, partial overlap}. By training a multi-class Gaussian Process Classifier (GPC) with these vectors on artificially perturbed data,
the goal is to classify ϕk
in three semantic uncertainty
concepts
C = {sensor noise, pose error, partial overlap}. At test time the GPC yields calibrated posteriors
πk(c) = p(c |ϕk
,D), c ∈ C,uk = −
X
c∈C
πk(c)logπk(c),
where uk summarizes overall predictive uncertainty. At test time the GPC yields calibrated posteriors
πk(c) = p(c |ϕk
,D), c ∈ C,uk = −
X
c∈C
πk(c)logπk(c),
where uk summarizes overall predictive uncertainty. In practice, πk(c) is computed via a standard multi-class GP link
with a Laplace approximation or Monte Carlo samples of the
latent scores. In practice, πk(c) is computed via a standard multi-class GP link
with a Laplace approximation or Monte Carlo samples of the
latent scores. Sampling from the GPC posterior provides both
per-concept attribution scores {πk(c)}c∈C and calibrated
uncertainty estimates uk. Sampling from the GPC posterior provides both
per-concept attribution scores {πk(c)}c∈C and calibrated
uncertainty estimates uk. This analysis inspired by TCAV
in [22] in the encoder space provides
sc(ek) = ∇e gc(e)
⊤
e=ek
vc,
where gc is the pre-link class score and vc is the concept
activation vector for concept c. This analysis inspired by TCAV
in [22] in the encoder space provides
sc(ek) = ∇e gc(e)
⊤
e=ek
vc,
where gc is the pre-link class score and vc is the concept
activation vector for concept c. This enables the system both
to measure how uncertain a registration is (uk) and to explain
why (via the dominant concept c
⋆
k = argmaxc πk(c)).Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 3 ---
Fig. 2: Schematic description of GUARD. p. 2 2: Schematic description of GUARD. 3: Overview of GUARD: concept-attributed perception drives
risk-adaptive RMPCC in real time. 2: Schematic description of GUARD. 3: Overview of GUARD: concept-attributed perception drives
risk-adaptive RMPCC in real time.Where the paper says it
Where the paper says it
Reactive Model Predictive Contouring Control (RMPCC Our prior work [4] presents RMPCC shown in Figure 1 that follows a reference path with CBFs. p. 1 Assuming a perfectly accurate perception, RMPCC proposes a
controller that considers CBF constraints. Reactive Model Predictive Contouring Control (RMPCC)
Our prior work [4] presents RMPCC shown in Figure 1
that follows a reference path with CBFs. Reactive Model Predictive Contouring Control (RMPCC)
Our prior work [4] presents RMPCC shown in Figure 1
that follows a reference path with CBFs. The robot path
is parameterized by a path parameter s ∈ [0,1], formalized
via a position spline pr(s) ∈ R
3
and an orientation curve
Rr(s) ∈ SO(3). The robot path
is parameterized by a path parameter s ∈ [0,1], formalized
via a position spline pr(s) ∈ R
3
and an orientation curve
Rr(s) ∈ SO(3). For the robot joint configuration q ∈ R
n,
the end-effector pose pee(q) and Ree(q), RMPCC refers to a
surrogate path–progress state (s,vs) with discrete dynamics. For the robot joint configuration q ∈ R
n,
the end-effector pose pee(q) and Ree(q), RMPCC refers to a
surrogate path–progress state (s,vs) with discrete dynamics. For the robot state and control input
xk =
q
⊤
k
sk vs,k⊤
, uk =
q˙
⊤
k
v˙s,k⊤
,
the controller minimizes tracking error and control effort over
a horizon N:
J = Lf (xN ) +
N
X−1
k=1
L(xk,uk,uk−1), (1)
with stage cost
L = wc∥ec∥
2 + wl∥el∥
2 + wvs
(vdes − vs,k)
2
| {z }
tracking
+wo∥Log(R
⊤
r Ree)∥
2
| {z }
orientation
+wq˙∥q˙k∥
2 + w∆ ˙q∥∆ ˙qk∥
2 + wv˙s
v˙
2
s,k
| {z }
regularization
, (2)
Here, e = pr(sk) − pee(qk) is split into contouring and
lag error using the local path tangent, ∆ ˙qk = ˙qk − q˙k−1
(with ∆ ˙q1 = 0), and Log(·) denotes the matrix logarithm on
SO(3). For the robot joint configuration q ∈ R
n,
the end-effector pose pee(q) and Ree(q), RMPCC refers to a
surrogate path–progress state (s,vs) with discrete dynamics. For the robot state and control input
xk =
q
⊤
k
sk vs,k⊤
, uk =
q˙
⊤
k
v˙s,k⊤
,
the controller minimizes tracking error and control effort over
a horizon N:
J = Lf (xN ) +
N
X−1
k=1
L(xk,uk,uk−1), (1)
with stage cost
L = wc∥ec∥
2 + wl∥el∥
2 + wvs
(vdes − vs,k)
2
| {z }
tracking
+wo∥Log(R
⊤
r Ree)∥
2
| {z }
orientation
+wq˙∥q˙k∥
2 + w∆ ˙q∥∆ ˙qk∥
2 + wv˙s
v˙
2
s,k
| {z }
regularization
, (2)
Here, e = pr(sk) − pee(qk) is split into contouring and
lag error using the local path tangent, ∆ ˙qk = ˙qk − q˙k−1
(with ∆ ˙q1 = 0), and Log(·) denotes the matrix logarithm on
SO(3). For the robot state and control input
xk =
q
⊤
k
sk vs,k⊤
, uk =
q˙
⊤
k
v˙s,k⊤
,
the controller minimizes tracking error and control effort over
a horizon N:
J = Lf (xN ) +
N
X−1
k=1
L(xk,uk,uk−1), (1)
with stage cost
L = wc∥ec∥
2 + wl∥el∥
2 + wvs
(vdes − vs,k)
2
| {z }
tracking
+wo∥Log(R
⊤
r Ree)∥
2
| {z }
orientation
+wq˙∥q˙k∥
2 + w∆ ˙q∥∆ ˙qk∥
2 + wv˙s
v˙
2
s,k
| {z }
regularization
, (2)
Here, e = pr(sk) − pee(qk) is split into contouring and
lag error using the local path tangent, ∆ ˙qk = ˙qk − q˙k−1
(with ∆ ˙q1 = 0), and Log(·) denotes the matrix logarithm on
SO(3). Safety is considered via CBFs on (i) kinematic singularities, (ii) self–collision, and (iii) environment collision. Safety is considered via CBFs on (i) kinematic singularities, (ii) self–collision, and (iii) environment collision. The algorithm proposes barrier functions h(q) ≥ 0
hsing(q) = µ(q) − εsing, (3)
µ(q) = q
detJee(q)Jee(q)⊤
, (4)
hself(q) = dself(q) − εself. 3: Overview of GUARD: concept-attributed perception drives
risk-adaptive RMPCC in real time. and
henv,ℓ(q, pobs) = denv,ℓ(q, pobs) − robs − εenv
for each link ℓ. and
henv,ℓ(q, pobs) = denv,ℓ(q, pobs) − robs − εenv
for each link ℓ. Here, Jee denotes the end-effector Jacobian;
dself and denv,ℓ are the minimum distances to self-collision
and to the ℓ-th robot link’s collision with the obstacle,
respectively, as predicted by a deep neural network; and
ε· > 0 denote safety margins. Here, Jee denotes the end-effector Jacobian;
dself and denv,ℓ are the minimum distances to self-collision
and to the ℓ-th robot link’s collision with the obstacle,
respectively, as predicted by a deep neural network; and
ε· > 0 denote safety margins. The optimization is solved
via QP relaxation of the barrier inequalities via Jacobian
linearization. The optimization is solved
via QP relaxation of the barrier inequalities via Jacobian
linearization. The Gauss–Newton Hessian approximation
enables real-time at 100 Hz. The Gauss–Newton Hessian approximation
enables real-time at 100 Hz. However, the method relies
on a perfectly accurate perception. However, the method relies
on a perfectly accurate perception. GUARD addresses this
limitation via an uncertainty-aware ICP.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 4 ---
Fig. 5: 7-DoF manipulator follows a 3D lemniscate with a dynamic obstacle; (a) setup; (b) minimum robot–obstacle distance for RMPCC and GUARD (blue) and red; dashed: safety margin εobs; green boxes indicate collisions. p. 3 distance
Fig. 5: 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. distance
Fig. 5: 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.Where the paper says it
Where the paper says it
--- Page 5 ---
Fig. 5: 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. p. 3 distance
Fig. 5: 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.Where the paper says it
--- Page 5 ---
Fig. 5: 7-DoF manipulator follows a 3D lemniscate with a dynamic obstacle; (a) setup; (b) minimum robot–obstacle distance for RMPCC and GUARD (blue) and red; dashed: safety margin εobs; green boxes indicate collisions.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Active PKO via Concept Feedback allows GUARD to adapt kernel parameters online by matching inlier distributions with empirical residuals, specifically adapting
the robust kernel parameters by matching an inlier distribution with the empirical residual
as described in Section II. -
Uncertainty-aware Robot Control enables the robot controller to modulate safety based on perception uncertainty, where "The scalar risk and per-concept posteriors modulate (i)the CBF relaxation (large slack penalties at high risk), (ii)collision margins and desired speed vdes, and (iii)soft weights wc,wl,wo,wvs to trade efficiency for conservatism if perception uncertainty spikes."
-
Risk Mapping and Safety under Parameter Adaptation allows GUARD to derive a bounded risk measure rho k from concept posteriors and entropy uk using the formula
rho k = clip[0,1] beta0 uk + X c∈C beta c πk(c),
which then linearly inflates margins, increases slack penalty, and reduces speed to tighten the safe set. -
The system can flag out-of-distribution frames using a reject rule defined as "OODk = 1 if uk > τu or max c∈C πk(c) < τp i,
which triggers freezing kernel updates and reducing v eff des while inflating margins via
(9)." -
GUARD combines traditional control with learning algorithms to foster the development of safe, yet efficient and flexible applications by unifying reactive model predictive contouring control (RMPCC) with uncertainty-aware ICP.
Abstract
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.
Sources
- Reactive Model Predictive Contouring Control for Robot Manipulators
- On the Covariance of ICP-based Scan-matching Techniques
- Stein ICP for Uncertainty Estimation in Point Cloud Matching
- Dynamic Graph CNN for Learning on Point Clouds
- Towards Explaining Uncertainty Estimates in Point Cloud Registration
Related papers
- Rethinking Actions: Continuity-Aware Priors for World-Model Planning
- MOSAIC: Modular Obstacle and Spatial-Aware Framework for Manipulation In Cluttered Scenes
- SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning
- DexSeed: Expanding Human Hand-Object Interactions into Robot Demonstrations for Dexterous Policy Learning
- LaST0.5: A Generalist World Action Model with Interaction-Centric Latent Reasoning
- CIALLO: Learning Shared Intent Representations for Cross-Embodiment Manipulation