pith. sign in

arxiv: 2605.21811 · v1 · pith:EWASAKXQnew · submitted 2026-05-20 · 💻 cs.RO

Safe and Steerable Geometric Motion Policies for Robotic Dexterous Manipulation

Albert Wu , Riccardo Bonalli , Thomas Lew , C. Karen Liu This is my paper

Pith reviewed 2026-05-22 08:30 UTC · model grok-4.3

classification 💻 cs.RO
keywords SafePBDSdexterous manipulationcontrol barrier functionsgeometric motion policiesrobotic graspingin-hand reorientationconfiguration manifoldsafety guarantees
0
0 comments X

The pith

SafePBDS computes optimal, certifiably safe configuration manifold accelerations from objectives and safety requirements on arbitrary task manifolds.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Robots must reconcile objectives and constraints defined on heterogeneous geometric spaces, such as tracking end-effector poses on SE(3) while avoiding obstacles in R^3, all while being controlled on a configuration manifold like R^7. SafePBDS combines predefined task manifold dynamical systems into autonomous motion and adds mechanisms to enforce safety and allow steering. Its pullback control barrier function construction turns task manifold safety conditions into linear constraints on configuration manifold accelerations. An action interface then lets high-level policies inject low-dimensional residual motions that preserve safety, so zero input recovers the original autonomous behavior. Experiments on a 23-DOF Franka Panda-Allegro Hand show 92.5% grasping success across 20 objects and the first model-based fully actuated palm-down in-hand reorientation exceeding 360 degrees of yaw.

Core claim

Safe Pullback Bundle Dynamical Systems (SafePBDS) is a geometrically consistent framework that computes optimal, certifiably safe configuration manifold accelerations from objectives and safety requirements on arbitrary task manifolds. It extends bundle dynamical systems with a pullback control barrier function construction that converts task manifold safety conditions into linear constraints on configuration manifold accelerations, and with a task manifold action interface that allows high-level policies to inject low-dimensional residual motions while preserving safety under arbitrary inputs.

What carries the argument

Pullback control barrier function construction, which converts safety conditions on task manifolds into linear constraints on configuration manifold accelerations while preserving safety certificates under bundle dynamical system dynamics.

If this is right

  • 92.5% success rate in dexterous grasping across 20 household objects and 120 trials
  • Exclusion of any one of four fingers via a one-dimensional action achieves 94.4% 3-finger grasp success
  • First model-based fully actuated palm-down in-hand reorientation exceeding 360 degrees of yaw rotation in both directions

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • High-level policies could optimize residual actions for sequences of manipulation tasks while retaining safety
  • The linear constraint structure may support real-time replanning when task manifolds are updated by perception
  • Similar pullback constructions could certify safety for mobile bases or dual-arm systems with heterogeneous task spaces

Load-bearing premise

Safety conditions defined on task manifolds can be converted into linear constraints on configuration manifold accelerations while preserving certificates of safety under the bundle dynamical system dynamics.

What would settle it

A robot motion that violates a specified task manifold safety condition, such as an obstacle avoidance margin in R^3, even though the configuration manifold accelerations were computed by SafePBDS.

Figures

Figures reproduced from arXiv: 2605.21811 by Albert Wu, C. Karen Liu, Riccardo Bonalli, Thomas Lew.

Figure 1
Figure 1. Figure 1: Overview of SafePBDS. At each control step, tasks on heterogeneous manifolds (Ni, gi) are pulled back to the configuration manifold M and composed into a safe acceleration a¯ via quadratic programs. (Left) Repre￾sentative tasks include an autonomous joint-damping SMCS (blue), a high￾level finger-position action (red), and a force-closure CBF safety constraint (green), each connected to M by a task map. (Ri… view at source ↗
Figure 2
Figure 2. Figure 2: Pullback CBF and action interface on S 2 ; run indices match Table II and the legend. (a) Autonomous runs (i)–(vii) on the same scene; recovery runs (vi)–(vii) start inside the obstacle (× marker). (b) h0(t) (top, shaded region is h0 < 0) and geodesic distance to goal (bottom) for the same runs. (c) (viii)–(xii): tangential actions ±u⊥ select opposite homotopy classes, uunsafe is clipped by the CBF, and (x… view at source ↗
Figure 3
Figure 3. Figure 3: 7-DOF arm: workspace safety and steered action. (a) Obstacle [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Hardware setup for the dexterous manipulation experiments: a 7-DOF [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: 4-finger grasp examples across representative object categories. All [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: 3-finger grasping variants for the three test objects. Each image is [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Per-(object, pose) grasp outcomes plotted against object weight and [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Offline IHR motion plan trees for (a) clockwise and (b) counter [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: In-hand reorientation on hardware. Snapshots are arranged temporally from left to right.(a) Clockwise rotation with the arm held static and empty [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗
read the original abstract

Robotic dexterous manipulation requires continuously reconciling objectives and constraints defined on heterogeneous geometric spaces: a robot controlled on a $\mathbb{R}^7$ configuration manifold may need to track end effector poses on $\mathrm{SE}(3)$ while satisfying obstacle avoidance margins in $\mathbb{R}$. We present Safe Pullback Bundle Dynamical Systems (SafePBDS), a geometrically consistent framework that computes optimal, certifiably safe configuration manifold accelerations from objectives and safety requirements on arbitrary task manifolds. SafePBDS builds on prior work that combines predefined task manifold dynamical systems to produce autonomous motion. Its first innovation is a pullback control barrier function construction, which converts task manifold safety conditions into linear constraints on configuration manifold accelerations. The second innovation is a task manifold action interface that allows a high-level policy to inject low dimensional residual motions; zero input recovers the autonomous behavior, while safety is preserved under arbitrary inputs. This lets high-level policies efficiently steer exploration while leaving precise motion to the autonomous behavior. We validate SafePBDS in simulation and on a 23-DOF Franka Panda-Allegro Hand platform. On dexterous grasping, SafePBDS achieves a $92.5\%$ success rate across 20 household objects and 120 trials. Using the action interface, the method can exclude any one of the four fingers during grasping via a one-dimensional action, achieving $94.4\%$ 3-finger grasp success across 3 objects and 36 trials. The efficient planning and safety guarantee of SafePBDS also enables the first model-based, fully actuated palm-down in-hand reorientation, exceeding $360^\circ$ of yaw rotation in both directions under varying object weight and wrist motion. Demo video and details: https://tml.stanford.edu/safe-pbds

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper presents Safe Pullback Bundle Dynamical Systems (SafePBDS), a geometrically consistent framework for dexterous manipulation that computes optimal, certifiably safe accelerations on the robot configuration manifold (e.g., R^7) from objectives and safety requirements defined on arbitrary task manifolds (e.g., SE(3) or R^3). It extends prior autonomous motion generation by introducing a pullback control barrier function (CBF) construction that converts task-manifold safety conditions into linear constraints on configuration accelerations, plus a task-manifold action interface that permits low-dimensional residual inputs from a high-level policy while preserving safety. Validation includes 92.5% grasping success across 120 trials on 20 objects with a 23-DOF Franka Panda-Allegro Hand, 94.4% 3-finger grasp success, and the first model-based fully actuated palm-down in-hand reorientation exceeding 360° yaw under varying conditions.

Significance. If the safety certificates are rigorously established, the work would offer a principled geometric approach to safe, steerable manipulation that reconciles heterogeneous manifolds without sacrificing autonomy or requiring heavy online optimization. The empirical results on high-DOF hardware, including exclusion of individual fingers via 1D actions and full reorientation, demonstrate practical utility. The framework's ability to recover autonomous behavior at zero input while guaranteeing safety under arbitrary steering inputs is a notable strength for integrating with learned high-level policies.

major comments (2)
  1. [§4] §4 (Pullback CBF Construction): The central claim that task-manifold safety conditions convert to linear constraints on configuration accelerations while preserving certificates under the bundle dynamical system is load-bearing for the 'certifiably safe' assertion. The skeptic note highlights that for curved task manifolds (e.g., SE(3) or S^2) the pullback of the barrier gradient may introduce higher-order curvature or holonomy terms; if these are approximated or dropped to enforce linearity, the Lie derivative condition for forward invariance may not hold along closed-loop trajectories. Please provide the explicit derivation (including any neglected terms) showing how the certificate remains valid for the full 23-DOF system.
  2. [§5.2] §5.2 and Table 1 (Experimental Validation): The reported 92.5% grasping success and 360° reorientation results are post-hoc across 120 and 36 trials respectively, but lack an ablation isolating the effect of the pullback approximation versus the action interface. Without this, it is difficult to assess whether the safety guarantee contributes to the observed performance or if failures are due to unmodeled dynamics rather than constraint violation.
minor comments (2)
  1. [§3] The notation for the bundle map and its differential in the pullback operation could be clarified with an explicit diagram or coordinate chart example, as the transition from task-manifold barrier to configuration-space linear constraint is central to reproducibility.
  2. [Figure 3] Figure 3 (or equivalent hardware setup figure): The caption should explicitly state the object weights and wrist motion ranges used in the reorientation experiments to allow direct comparison with future work.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thoughtful and constructive review of our manuscript. We address each major comment below, providing clarifications on the theoretical construction and experimental analysis. We indicate revisions that will be incorporated in the next version of the paper.

read point-by-point responses

  1. Referee: [§4] §4 (Pullback CBF Construction): The central claim that task-manifold safety conditions convert to linear constraints on configuration accelerations while preserving certificates under the bundle dynamical system is load-bearing for the 'certifiably safe' assertion. The skeptic note highlights that for curved task manifolds (e.g., SE(3) or S^2) the pullback of the barrier gradient may introduce higher-order curvature or holonomy terms; if these are approximated or dropped to enforce linearity, the Lie derivative condition for forward invariance may not hold along closed-loop trajectories. Please provide the explicit derivation (including any neglected terms) showing how the certificate remains valid for the full 23-DOF system.

    Authors: We thank the referee for this important observation on the rigor of the safety certificate. The pullback CBF is constructed by composing the task-manifold barrier h with the forward map φ: Q → M, yielding the pulled-back barrier h ∘ φ on the configuration manifold Q. Its gradient is obtained via the adjoint of the differential: ∇(h ∘ φ) = Dφ^* ∇h. The first Lie derivative along the velocity is L_f (h ∘ φ) = ⟨∇(h ∘ φ), f⟩. The second derivative, which supplies the linear constraint on configuration acceleration, follows from the chain rule applied to the bundle dynamical system; curvature and connection terms arising from the geometry of M appear explicitly in the Hessian contribution and are retained in the expression. No terms are dropped to enforce linearity—the linearity in acceleration is a direct consequence of the second-order CBF condition. The bundle structure ensures that forward invariance of the safe set on M is preserved on Q. We will expand Section 4 with the complete derivation, including all curvature terms, and verify the closed-loop certificate for the 23-DOF system in the revised manuscript. revision: yes

  2. Referee: [§5.2] §5.2 and Table 1 (Experimental Validation): The reported 92.5% grasping success and 360° reorientation results are post-hoc across 120 and 36 trials respectively, but lack an ablation isolating the effect of the pullback approximation versus the action interface. Without this, it is difficult to assess whether the safety guarantee contributes to the observed performance or if failures are due to unmodeled dynamics rather than constraint violation.

    Authors: We agree that additional analysis would help readers evaluate the contributions. The safety guarantee is established by construction in Theorem 1 and holds for arbitrary inputs through the action interface; the pullback operation is an exact geometric construction rather than an approximation. In the reported trials we continuously monitored the value of the pulled-back barrier function and observed that it remained strictly positive, indicating that constraint violations did not occur. Observed failures are attributable to unmodeled contact dynamics, perception noise, or object properties outside the modeled friction and mass ranges. To address the referee’s concern we will add a dedicated paragraph in Section 5.2 that (i) recalls the theoretical separation between the CBF guarantee and the action interface, (ii) reports the barrier-function traces from the hardware experiments, and (iii) categorizes the failure modes with supporting data. A controlled ablation isolating every modeling choice would require new hardware trials; we therefore provide the requested discussion and monitoring evidence instead. revision: partial

Circularity Check

0 steps flagged

No significant circularity detected; independent safety and steering constructions

full rationale

The derivation chain for SafePBDS starts from prior autonomous motion generation on task manifolds and adds two explicitly new elements: a pullback control barrier function that converts task-manifold safety conditions into linear constraints on configuration accelerations, and a task-manifold action interface that injects residual motions while preserving safety. Neither construction is shown to reduce by definition or by self-citation to the inputs of the other; the abstract presents them as innovations that extend the base framework without circular re-use of fitted parameters or uniqueness theorems. Experimental results on the 23-DOF hand supply external validation rather than internal fitting, confirming the central claims remain self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

Abstract provides limited visibility into internal assumptions; relies on geometric consistency of pullback operations and predefined task dynamical systems.

axioms (1)
  • domain assumption Predefined task manifold dynamical systems can be combined to produce autonomous motion on the configuration manifold.
    Framework builds directly on prior work that combines such systems.
invented entities (2)
  • pullback control barrier function no independent evidence
    purpose: Converts task manifold safety conditions into linear constraints on configuration manifold accelerations.
    Described as the first innovation enabling certifiable safety.
  • task manifold action interface no independent evidence
    purpose: Allows high-level policy to inject low-dimensional residual motions while preserving safety.
    Described as the second innovation for steerability.

pith-pipeline@v0.9.0 · 5858 in / 1341 out tokens · 55470 ms · 2026-05-22T08:30:08.474769+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

78 extracted references · 78 canonical work pages · 4 internal anchors

  1. [1]

    Control Barrier Functions: The- ory and Applications

    Aaron D. Ames et al. “Control Barrier Functions: The- ory and Applications”. In:2019 18th European Control Conference (ECC). IEEE, 2019, pp. 3420–3431

    work page 2019

  2. [2]

    Surprisingly Robust In-Hand Ma- nipulation: An Empirical Study

    Aditya Bhatt et al. “Surprisingly Robust In-Hand Ma- nipulation: An Empirical Study”. In:Proceedings of Robotics: Science and Systems (RSS). 2021.DOI: 10. 15607/RSS.2021.XVII.089

    work page 2021

  3. [3]

    Data-Driven Grasp Synthesis— A Survey

    Jeannette Bohg et al. “Data-Driven Grasp Synthesis— A Survey”. In:IEEE Transactions on Robotics30.2 (2014), pp. 289–309. (a) Clockwise rotation, static arm (b) Counterclockwise rotation, swinging arm with weighted object Fig. 9. In-hand reorientation on hardware. Snapshots are arranged temporally from left to right.(a) Clockwise rotation with the arm held...

    work page 2014

  4. [4]

    Lewis.Geometric Con- trol of Mechanical Systems

    Francesco Bullo and Andrew D. Lewis.Geometric Con- trol of Mechanical Systems. V ol. 49. Texts in Applied Mathematics. New York-Heidelberg-Berlin: Springer Verlag, 2004.ISBN: 0-387-22195-6

    work page 2004

  5. [6]

    A System for General In-Hand Object Re-Orientation

    Tao Chen, Jie Xu, and Pulkit Agrawal. “A System for General In-Hand Object Re-Orientation”. In:Proceed- ings of the 5th Conference on Robot Learning (CoRL). V ol. 164. Proceedings of Machine Learning Research. PMLR, 2022, pp. 297–307

    work page 2022

  6. [7]

    Visual Dexterity: In-Hand Reorienta- tion of Novel and Complex Object Shapes

    Tao Chen et al. “Visual Dexterity: In-Hand Reorienta- tion of Novel and Complex Object Shapes”. In:Science Robotics8.84 (2023), eadc9244

    work page 2023

  7. [8]

    RMPflow: A Geometric Frame- work for Generation of Multitask Motion Policies

    Ching-An Cheng et al. “RMPflow: A Geometric Frame- work for Generation of Multitask Motion Policies”. In: IEEE Transactions on Automation Science and Engi- neering18.3 (2021), pp. 968–987

    work page 2021

  8. [9]

    Diffusion Policy: Visuomotor Pol- icy Learning via Action Diffusion

    Cheng Chi et al. “Diffusion Policy: Visuomotor Pol- icy Learning via Action Diffusion”. In:The Interna- tional Journal of Robotics Research44.10-11 (2025), pp. 1684–1704

    work page 2025

  9. [10]

    A Report on Harmonic Maps

    James Eells and Luc Lemaire. “A Report on Harmonic Maps”. In:Bulletin of the London Mathematical Society 10.1 (1978), pp. 1–68

    work page 1978

  10. [11]

    The GRASP Taxonomy of Hu- man Grasp Types

    Thomas Feix et al. “The GRASP Taxonomy of Hu- man Grasp Types”. In:IEEE Transactions on Human- Machine Systems46.1 (2016), pp. 66–77

    work page 2016

  11. [12]

    Gemini Robotics Team, Google DeepMind.Gemini Robotics 1.5: Pushing the Frontier of Generalist Robots with Advanced Embodied Reasoning, Thinking, and Motion Transfer. 2025. arXiv: 2510.03342[cs.RO]

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  12. [13]

    DeXtreme: Transfer of Agile In- Hand Manipulation from Simulation to Reality

    Ankur Handa et al. “DeXtreme: Transfer of Agile In- Hand Manipulation from Simulation to Reality”. In: Proceedings of the IEEE International Conference on Robotics and Automation (ICRA). 2023, pp. 5977–5984

    work page 2023

  13. [14]

    A Unified Approach for Motion and Force Control of Robot Manipulators: The Operational Space Formulation

    Oussama Khatib. “A Unified Approach for Motion and Force Control of Robot Manipulators: The Operational Space Formulation”. In:IEEE Journal on Robotics and Automation3.1 (1987), pp. 43–53

    work page 1987

  14. [17]

    Lee.Introduction to Smooth Manifolds

    John M. Lee.Introduction to Smooth Manifolds. 2nd ed. V ol. 218. Graduate Texts in Mathematics. Springer, 2012

    work page 2012

  15. [18]

    Survey, Georgia Institute of Technology

    Hrishit Leen et al.Deep Learning for Dexterous Robot Grasping. Survey, Georgia Institute of Technology. 2025

    work page 2025

  16. [20]

    Review on Motion Planning of Robotic Manipu- lator in Dynamic Environments

    Jie Liu, Hwa Jen Yap, and Anis Salwa Mohd Khairud- din. “Review on Motion Planning of Robotic Manipu- lator in Dynamic Environments”. In:Journal of Sensors 2024 (2024), p. 5969512

    work page 2024

  17. [21]

    Synthesizing Diverse and Physi- cally Stable Grasps With Arbitrary Hand Structures Us- ing Differentiable Force Closure Estimator

    Tengyu Liu et al. “Synthesizing Diverse and Physi- cally Stable Grasps With Arbitrary Hand Structures Us- ing Differentiable Force Closure Estimator”. In:IEEE Robotics and Automation Letters7.1 (2022), pp. 470– 477

    work page 2022

  18. [22]

    DextrAH-G: Pixels-to-Action Dexterous Arm-Hand Grasping with Geometric Fab- rics

    Tyler Ga Wei Lum et al. “DextrAH-G: Pixels-to-Action Dexterous Arm-Hand Grasping with Geometric Fab- rics”. In:Proceedings of The 8th Conference on Robot Learning (CoRL). V ol. 270. Proceedings of Machine Learning Research. PMLR, 2025, pp. 3182–3211

    work page 2025

  19. [23]

    Complex In-Hand Manipula- tion via Compliance-Enabled Finger Gaiting and Multi- Modal Planning

    Andrew S. Morgan et al. “Complex In-Hand Manipula- tion via Compliance-Enabled Finger Gaiting and Multi- Modal Planning”. In:IEEE Robotics and Automation Letters7.2 (2022), pp. 4821–4828

    work page 2022

  20. [24]

    Murray, Zexiang Li, and S

    Richard M. Murray, Zexiang Li, and S. Shankar Sastry. A Mathematical Introduction to Robotic Manipulation. CRC Press, 1994

    work page 1994

  21. [25]

    Deep Learning Approaches to Grasp Synthesis: A Review

    Rhys Newbury et al. “Deep Learning Approaches to Grasp Synthesis: A Review”. In:IEEE Transactions on Robotics39.5 (2023), pp. 3994–4015

    work page 2023

  22. [26]

    Exponential Con- trol Barrier Functions for enforcing high relative-degree safety-critical constraints

    Quan Nguyen and Koushil Sreenath. “Exponential Con- trol Barrier Functions for enforcing high relative-degree safety-critical constraints”. In:2016 American Control Conference (ACC). IEEE. 2016, pp. 322–328

    work page 2016

  23. [27]

    Learning Dexterous In-Hand Manip- ulation

    OpenAI et al. “Learning Dexterous In-Hand Manip- ulation”. In:The International Journal of Robotics Research39.1 (2020), pp. 3–20

    work page 2020

  24. [28]

    Global Planning for Contact-Rich Manipulation via Local Smoothing of Quasi-Dynamic Contact Models

    Tao Pang et al. “Global Planning for Contact-Rich Manipulation via Local Smoothing of Quasi-Dynamic Contact Models”. In:IEEE Transactions on Robotics 39.6 (2023), pp. 4691–4711

    work page 2023

  25. [29]

    Physical Intelligence.π 0.7: A Steerable Generalist Robotic Foundation Model with Emergent Capabilities

    work page

  26. [30]

    arXiv: 2604.15483[cs.RO]

    work page internal anchor Pith review Pith/arXiv arXiv

  27. [31]

    Learning a Shape-Conditioned Agent for Purely Tactile In-Hand Manipulation of Vari- ous Objects

    Johannes Pitz et al. “Learning a Shape-Conditioned Agent for Purely Tactile In-Hand Manipulation of Vari- ous Objects”. In:2024 IEEE/RSJ International Confer- ence on Intelligent Robots and Systems (IROS). IEEE. 2024, pp. 13112–13119

    work page 2024

  28. [32]

    A direct method for trajectory optimization of rigid bod- ies through contact

    Michael Posa, Cecilia Cantu, and Russ Tedrake. “A direct method for trajectory optimization of rigid bod- ies through contact”. In:The International Journal of Robotics Research33.1 (2014), pp. 69–81

    work page 2014

  29. [33]

    General In-Hand Object Rotation with Vision and Touch

    Haozhi Qi et al. “General In-Hand Object Rotation with Vision and Touch”. In:Proceedings of The 7th Conference on Robot Learning (CoRL). V ol. 229. Pro- ceedings of Machine Learning Research. PMLR, 2023, pp. 2549–2564

    work page 2023

  30. [34]

    Riemannian Motion Policies

    Nathan D Ratliff et al. “Riemannian motion policies”. In:arXiv preprint arXiv:1801.02854(2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  31. [35]

    Optimization Fabrics

    Nathan D. Ratliff et al. “Optimization Fabrics”. In: arXiv preprint arXiv:2008.02399(2020). 17

    work page arXiv 2008

  32. [36]

    From Bundles to Backstepping: Geometric Control Barrier Functions for Safety-Critical Control on Mani- folds

    Massimiliano de Sa, Pio Ong, and Aaron D. Ames. “From Bundles to Backstepping: Geometric Control Barrier Functions for Safety-Critical Control on Mani- folds”. In:arXiv preprint arXiv:2510.20202(2025)

    work page arXiv 2025

  33. [37]

    Synthesis of Whole- Body Behaviors through Hierarchical Control of Behav- ioral Primitives

    Luis Sentis and Oussama Khatib. “Synthesis of Whole- Body Behaviors through Hierarchical Control of Behav- ioral Primitives”. In:International Journal of Humanoid Robotics2.4 (2005), pp. 505–518

    work page 2005

  34. [38]

    Control Barrier Func- tions for Mechanical Systems: Theory and Application to Robotic Grasping

    Wenceslao Shaw Cortez et al. “Control Barrier Func- tions for Mechanical Systems: Theory and Application to Robotic Grasping”. In:IEEE Transactions on Control Systems Technology29.2 (2021), pp. 530–545

    work page 2021

  35. [39]

    Robust pivoting: Exploiting frictional stability using bilevel optimization

    Yuki Shirai et al. “Robust pivoting: Exploiting frictional stability using bilevel optimization”. In:2022 Interna- tional Conference on Robotics and Automation (ICRA). IEEE. 2022, pp. 992–998

    work page 2022

  36. [40]

    A Gen- eral Framework for Managing Multiple Tasks in Highly Redundant Robotic Systems

    Bruno Siciliano and Jean-Jacques E. Slotine. “A Gen- eral Framework for Managing Multiple Tasks in Highly Redundant Robotic Systems”. In:Proceedings of the 5th International Conference on Advanced Robotics (ICAR). 1991, pp. 1211–1216

    work page 1991

  37. [41]

    Learning Purely Tactile In-Hand Manipulation with a Torque-Controlled Hand

    Leon Sievers, Johannes Pitz, and Berthold B ¨auml. “Learning Purely Tactile In-Hand Manipulation with a Torque-Controlled Hand”. In:2022 International Con- ference on Robotics and Automation (ICRA). IEEE. 2022, pp. 2745–2751

    work page 2022

  38. [42]

    Dexterous Contact-Rich Manip- ulation via the Contact Trust Region

    H. J. Terry Suh et al. “Dexterous Contact-Rich Manip- ulation via the Contact Trust Region”. In:The Interna- tional Journal of Robotics Research(2026). OnlineFirst. DOI: 10.1177/02783649251398875

    work page doi:10.1177/02783649251398875 2026

  39. [43]

    A survey of robot manipulation in contact

    Markku Suomalainen, Yiannis Karayiannidis, and Ville Kyrki. “A survey of robot manipulation in contact”. In:Robotics and Autonomous Systems156 (2022), p. 104224

    work page 2022

  40. [44]

    Deep Reinforcement Learning for Robotics: A Survey of Real-World Successes

    Chen Tang et al. “Deep Reinforcement Learning for Robotics: A Survey of Real-World Successes”. In: Annual Review of Control, Robotics, and Autonomous Systems8 (2025), pp. 153–188

    work page 2025

  41. [45]

    Safe Backstepping with Control Barrier Functions

    Andrew J. Taylor et al. “Safe Backstepping with Control Barrier Functions”. In:2022 IEEE 61st Conference on Decision and Control (CDC). IEEE. 2022, pp. 5775– 5782

    work page 2022

  42. [46]

    Mu- JoCo: A physics engine for model-based control

    Emanuel Todorov, Tom Erez, and Yuval Tassa. “Mu- JoCo: A physics engine for model-based control”. In: 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE. 2012, pp. 5026– 5033

    work page 2012

  43. [47]

    A Careful Examination of Large Behavior Models for Multitask Dexterous Manipulation

    TRI LBM Team et al. “A Careful Examination of Large Behavior Models for Multitask Dexterous Manipula- tion”. In: (2025). arXiv: 2507.05331[cs.RO]

    work page internal anchor Pith review arXiv 2025

  44. [48]

    A Survey on Imitation Learning for Contact-Rich Tasks in Robotics

    Toshiaki Tsuji et al. “A Survey on Imitation Learning for Contact-Rich Tasks in Robotics”. In:The Interna- tional Journal of Robotics Research(2026). OnlineFirst. DOI: 10.1177/02783649261417694

    work page doi:10.1177/02783649261417694 2026

  45. [49]

    Fast-Grasp’D: Dexterous Multi- finger Grasp Generation Through Differentiable Sim- ulation

    Dylan Turpin et al. “Fast-Grasp’D: Dexterous Multi- finger Grasp Generation Through Differentiable Sim- ulation”. In:2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE. 2023, pp. 8082–8089

    work page 2023

  46. [50]

    Geometric Fabrics: Generalizing Classical Mechanics to Capture the Physics of Behav- ior

    Karl Van Wyk et al. “Geometric Fabrics: Generalizing Classical Mechanics to Capture the Physics of Behav- ior”. In:IEEE Robotics and Automation Letters7.2 (2022), pp. 3202–3209

    work page 2022

  47. [51]

    Survey of Learning- Based Approaches for Robotic In-Hand Manipulation

    Abraham Itzhak Weinberg et al. “Survey of Learning- Based Approaches for Robotic In-Hand Manipulation”. In:Frontiers in Robotics and AI11 (2024), p. 1455431

    work page 2024

  48. [53]

    Learn- ing Diverse and Physically Feasible Dexterous Grasps with Generative Model and Bilevel Optimization

    Albert Wu, Michelle Guo, and C. Karen Liu. “Learn- ing Diverse and Physically Feasible Dexterous Grasps with Generative Model and Bilevel Optimization”. In: Proceedings of The 6th Conference on Robot Learning (CoRL). V ol. 205. Proceedings of Machine Learning Research. PMLR, 2023, pp. 1938–1948

    work page 2023

  49. [54]

    Safety-critical and constrained geometric control synthesis using control Lyapunov and control Barrier functions for systems evolving on manifolds

    Guofan Wu and Koushil Sreenath. “Safety-critical and constrained geometric control synthesis using control Lyapunov and control Barrier functions for systems evolving on manifolds”. In:2015 American Control Conference (ACC). IEEE. 2015, pp. 2038–2044

    work page 2015

  50. [55]

    High-Order Control Barrier Functions

    Wei Xiao and Calin Belta. “High-Order Control Barrier Functions”. In:IEEE Transactions on Automatic Con- trol67.7 (2022), pp. 3655–3662

    work page 2022

  51. [56]

    AnyRotate: Gravity-Invariant In-Hand Object Rotation with Sim-to-Real Touch

    Max Yang et al. “AnyRotate: Gravity-Invariant In-Hand Object Rotation with Sim-to-Real Touch”. In:Proceed- ings of The 8th Conference on Robot Learning (CoRL). V ol. 270. Proceedings of Machine Learning Research. PMLR, 2025, pp. 4727–4747

    work page 2025

  52. [57]

    Rotating without Seeing: To- wards In-hand Dexterity through Touch

    Zhao-Heng Yin et al. “Rotating without Seeing: To- wards In-hand Dexterity through Touch”. In:Proceed- ings of Robotics: Science and Systems (RSS). 2023.DOI: 10.15607/RSS.2023.XIX.036

    work page doi:10.15607/rss.2023.xix.036 2023

  53. [58]

    Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware

    Tony Z. Zhao et al. “Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware”. In:Proceed- ings of Robotics: Science and Systems (RSS). 2023.DOI: 10.15607/RSS.2023.XIX.016

    work page doi:10.15607/rss.2023.xix.016 2023

  54. [59]

    A Survey of Optimization-Based Task and Motion Planning: From Classical to Learning Approaches

    Zhigen Zhao et al. “A Survey of Optimization-Based Task and Motion Planning: From Classical to Learning Approaches”. In:IEEE/ASME Transactions on Mecha- tronics(2024). Early Access.DOI: 10.1109/TMECH. 2024.3452509

    work page doi:10.1109/tmech 2024

  55. [61]

    A multi-level optimization framework for simultaneous grasping and motion plan- ning

    Simon Zimmermann et al. “A multi-level optimization framework for simultaneous grasping and motion plan- ning”. In:IEEE Robotics and Automation Letters5.2 (2020), pp. 2966–2972. A1 APPENDIXA PROOF OFLEMMAIV.2 Lemma IV .2.In the task manifold PBDS setup, the BCBF candidatehgiven in(25)has time derivative ˙h=⟨gradh 0,˙x⟩N +ε ex,∇ ˙x˜ξ N −ε ex, d f M ∇ ˙σ˙σ ...

    work page 2020

  56. [62]

    safe velocity field

    The south pole chartφ S :U S →R 2 projects fromS= (0,0,−1): φS(x1, x2, x3) = x1 1 +x 3 , x2 1 +x 3 ,(55) with inverse ¯φS(y1, y2) = 2y1 ∥y∥2 + 1, 2y2 ∥y∥2 + 1, 1− ∥y∥ 2 ∥y∥2 + 1 .(56) The metric in both charts is given byg ij = 4 (1+∥y∥2)2 δij, with Christoffel symbols Γk ij = −2 1 +∥y∥ 2 (yiδjk +y jδik −y kδij).(57) A2 b) Configuration Dynamics:We compar...

    work page

  57. [63]

    Anorientation attractortask with scalar task map f:M→R,f(σ) =∥˜q ee(σ)−q goal∥2, the chord distance in theR 4 quaternion embedding between the goal and the end effector unit quaternion, with˜qee = sign(q⊤ eeqgoal)q ee resolving the antipodal±qambiguity in the universal double coverS 3 →SO(3). The task uses the quadratic potentialΦ(d) =d 2 and the Euclidea...

    work page

  58. [64]

    Aposition attractortask (when active) with task map f:M→R 3, with a quadratic potential attracting toward a goal position, with the Euclidean metric; a companion damping task applies linear dissipation with identity map and Euclidean metric onR 3

    work page

  59. [65]

    4)Joint limit ECBFtasks with task mapsf k :M→R, k= 1,

    Ajoint space dampingtask withf= id:M→R 7. 4)Joint limit ECBFtasks with task mapsf k :M→R, k= 1, . . . ,7, given by coordinate projections, each with a safety functionh 0,k enforced via the pullback ECBF (Section IV). 5)Workspace obstacle ECBFtasks (one per collision body) with per-body safety functionsh 0,j(σ) = dgeom(gj, g obs), whered geom is the signed...

    work page

  60. [66]

    Asteered actionwith scalar control task mapf:M→ R:σ7→σ 1, projecting onto joint 1, so the actionu∈R is a residual joint 1 acceleration. All experiments use the same shared scenario unless otherwise noted: the arm starts at its home configuration and must reach a goal end effector orientation obtained from the home orientation by an XYZ Euler rotation of(4...

    work page

  61. [67]

    For grasping, the damping matrix is diagonal with separate entries,B arm = 20I 7 andB hand = 0.1I 16

    Ajoint space dampingtask with task mapf D = id:M→R m, equipped with a dissipative forceF D = −B˙σand the flat metric. For grasping, the damping matrix is diagonal with separate entries,B arm = 20I 7 andB hand = 0.1I 16. For in-hand reorientation, the arm is fixed andB hand = 2.0I 16. 2)Joint limit ECBFtasks with task mapsf k :M→R,k= 1, . . . , m, given by...

    work page

  62. [68]

    The stable rest pose of each object is determined and permuted across 3 table locations

    work page

  63. [69]

    FoundationPose [7] estimates the object’s 6-DOF pose via the Redis streaming interface

    work page

  64. [70]

    For each yaw, a top-down pose (palm facing downward) is computed above the object’s bounding-box center

    Candidate wrist poses are generated by sampling 18 yaw angles (every20 ◦) around the vertical axis. For each yaw, a top-down pose (palm facing downward) is computed above the object’s bounding-box center. Three standoff distances (0, 7.5, 15 mm) along the approach direction are evaluated, yielding up to 54 candidates per object

    work page

  65. [71]

    Valid candidates are ranked by a quality metric that favors tight width fit and small standoff distance, and the top-ranked candidate is selected for hardware execution

    Each candidate is filtered through a five-stage pipeline: (i)object width check: the object extent along the thumb- to-index aperture axis must not exceed1.5times the open-hand span; (ii)arm inverse kinematics (IK): a MuJoCo-based IK solver verifies reachability with a 2 cm position tolerance,5 ◦ rotation tolerance, and a 5% joint limit margin; (iii)palm–...

    work page

  66. [72]

    Using the latest perception update, candidate generation and ranking are repeated at the start of each hardware attempt, so the chosen pose reflects the current object estimate

    work page

  67. [73]

    On hardware, the arm first moves to its home configu- ration and the hand opens to a pre-grasp pose (using a two-phase open: fingertip impedance control for coarse motion, then joint PD for fine positioning). The arm then executes a two-stage approach: first moving to a waypoint 15 cm above the grasp pose, then descending to the final wrist pose via joint...

    work page

  68. [74]

    next movable

    The simulation and hardware joint configurations are synchronized, after whichSafePBDStakes over the full23-DOF arm–hand system in MuJoCo [6]. Two action tasks are composed at each control step: (i) per- fingerfingertip-to-object distancetasks that PD-servo each fingertip toward the object surface with a clipped acceleration command, and (ii) anaverage-fi...

    work page

  69. [75]

    Step sizes ofδ∈ {10,20,30,40}mm on the world XY plane along the object surface, and

    work page

  70. [76]

    Aperture anglesϕ∈ {30 ◦,50 ◦,70 ◦,90 ◦}controlling the finger’s approach direction in the plane perpendicular to the object axis. 2 Lift-and-Drop State Machine Each candidate action is forward-simulated using SafePBDSand evaluated by a four-phase state machine that governs a single lift-and-drop primitive:LIFTING, TRAVERSING,DROPPING, andADJUSTING. InLIFT...

    work page

  71. [77]

    Composable Geometric Motion Policies using Multi- Task Pullback Bundle Dynamical Systems

    Andrew Bylard, Riccardo Bonalli, and Marco Pavone. “Composable Geometric Motion Policies using Multi- Task Pullback Bundle Dynamical Systems”. In:2021 IEEE International Conference on Robotics and Au- tomation (ICRA). IEEE. 2021, pp. 7464–7470

    work page 2021

  72. [78]

    https: //www.kiriengine.app/

    KIRI Innovations.KIRI Engine: 3D Scanner App. https: //www.kiriengine.app/. Photogrammetry and LiDAR- based 3D scanning application. 2024

    work page 2024

  73. [79]

    Segment Anything

    Alexander Kirillov et al. “Segment Anything”. In:Pro- ceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 2023, pp. 3992–4003

    work page 2023

  74. [80]

    Lee.Introduction to Riemannian Manifolds

    John M. Lee.Introduction to Riemannian Manifolds. 2nd ed. V ol. 176. Graduate Texts in Mathematics. Springer, 2018

    work page 2018

  75. [81]

    FRoGGeR: Fast Robust Grasp Gen- eration via the Min-Weight Metric

    Albert H. Li et al. “FRoGGeR: Fast Robust Grasp Gen- eration via the Min-Weight Metric”. In:Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). 2023, pp. 6809–6816

    work page 2023

  76. [82]

    Mu- JoCo: A physics engine for model-based control

    Emanuel Todorov, Tom Erez, and Yuval Tassa. “Mu- JoCo: A physics engine for model-based control”. In: 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE. 2012, pp. 5026– 5033. A7

    work page 2012

  77. [83]

    FoundationPose: Unified 6D Pose Estimation and Tracking of Novel Objects

    Bowen Wen et al. “FoundationPose: Unified 6D Pose Estimation and Tracking of Novel Objects”. In:Pro- ceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR). 2024, pp. 17868– 17879

    work page 2024

  78. [84]

    VIOLA: Imitation Learning for Vision-Based Manipulation with Object Proposal Pri- ors

    Yifeng Zhu et al. “VIOLA: Imitation Learning for Vision-Based Manipulation with Object Proposal Pri- ors”. In:Proceedings of The 6th Conference on Robot Learning (CoRL). V ol. 205. Proceedings of Machine Learning Research. PMLR, 2023, pp. 1199–1210

    work page 2023