PaperlayerView exact revision on arXiv

Source-linked AI summary

Optimize, Learn, Refine: Whole-Body Grasping and Pick-and-Throw with a Spiral Soft Robot

Marwah Basuhai, Tingcong Liu, Ibrahim Alsarraj, Yuhao Wang, Ke Wu

arXiv:2609.38202v2cs.RO

TL;DR

The paper addresses the difficulty of synthesizing reliable whole-body grasping and throwing behavior as contacts change, including acquisition from an initially ungrasped state. It uses outcome-based actuation optimization with learned warm starts and refinement, achieving high success in simulation and hardware.

  • Problem

    Changing contacts couple tendon actuation, body deformation, friction, and object motion, while independent search is costly and direct predicted commands may not reproduce outcomes under new conditions.

    Source: §I · p. 1

  • Method

    An outcome-based optimize–learn–refine framework uses CMA–ES to optimize tendon commands and a neural model to reuse prior solutions as warm starts for condition-specific refinement.

    Source: §I · p. 1

  • Results

    98.4% of randomized simulated grasps succeeded, throwing success was 98%, 97%, and 94% across three directions, and hardware success was 100% for both tasks.

    Source: §I · p. 1, C. Simulation Results · p. 6

  • Takeaways & Limitations

    The reported simulation and hardware results demonstrate the framework’s effectiveness across simulated and physical whole-body grasping and pick-and-throw tasks.

    Source: §VI · p. 8

Abstract

from arXiv · show

Soft continuum robots can exploit distributed compliance for whole-body manipulation, but synthesizing behavior through changing contacts remains difficult. We address whole-body grasping and pick-and-throw from an initially ungrasped state through outcome-based actuation-space optimization. Grasping is quantified by tip angular sweep and body-object enclosure, while throwing further incorporates release-direction alignment and minimum release speed. These objectives allow grasping, acceleration, and release to emerge from compliant interaction without prescribing contact forces, contact locations, or body configurations. Because the resulting actuation-to-outcome mapping is nonsmooth, we utilize derivative-free CMA-ES within an optimize-learn-refine framework. CMA-ES generates solutions for sampled conditions, a task-conditioned predictor learns warm starts, and CMA-ES refines them for unseen conditions. In simulation, the method achieves 492/500 successful grasps (98.4%) and success rates of 98%, 97%, and 94% across three directional throwing trials. Learned initialization increases grasping success from 78.6% to 98.4% while reducing the median rollout count from 1184 to 816 in CMA-ES. Hardware experiments achieve a 100% grasping success rate across 50 executions and a 100% pick-and-throw success rate across 30 executions, with 10 repetitions per direction. Together, these simulation and hardware results demonstrate the effectiveness of the proposed framework across both simulated and physical whole-body manipulation tasks.

I. INTRODUCTION

Soft continuum robots use distributed compliance for whole-body manipulation, but dynamic pick-and-throw must coordinate acquisition, retention, acceleration, and release as contacts change. The paper introduces an outcome-based optimize–learn–refine framework and evaluates it in simulation and on hardware.

I. INTRODUCTION

  • Prior throwing setups use separate grippers, end-effector attachments, or prescribed initial grasps, rather than integrating acquisition through release from an initially ungrasped state.Changing contacts couple tendon actuation, body deformation, friction, and object motion, making the actuation-to-outcome mapping potentially nonsmooth.Source: §I · p. 1
  • Grasping is formulated as terminal whole-body enclosure and extended to throwing with release-direction and minimum-speed requirements.These outcome requirements do not prescribe contact forces, locations, or body configurations.Source: §I · p. 1
  • CMA–ES optimizes tendon commands through rollouts, while a neural model reuses prior solutions to initialize refinement for new conditions.This combines cross-condition behavior reuse with condition-specific actuation search.Source: §I · p. 1
  • 98.4% simulated grasp success and 98%, 97%, and 94% throwing success across three directions accompany 100% hardware success in both tasks.Hardware evaluation comprised 50 grasping executions and 30 pick-and-throw executions, with 10 repetitions per direction.Source: §I · p. 1

II. PROBLEM STATEMENT

  • The problem statement defines the grasping and pick-and-throw tasks and introduces the tendon-driven soft robot SpiRob.Source: §II · p. 2

A. The Studied Tendon-Driven Soft Robot: SpiRob

SpiRob is a spiral-inspired tendon-driven soft robot with a backbone that grades from a relatively rigid base to a compliant tip. The paper considers planar grasping and pick-and-throw from initially ungrasped states, with tendon-command trajectories optimized through complete simulation rollouts.

A. The Studied Tendon-Driven Soft Robot: SpiRob

  • SpiRob has a logarithmic-spiral-inspired backbone whose curvature and stiffness vary continuously from a relatively rigid base to a highly compliant tip.The graded structure supports a large reachable workspace and reaching, wrapping, and grasping motions.Source: A. The Studied Tendon-Driven Soft… · p. 2

B. Manipulation Task Objectives

  • Both planar tasks start from a specified, initially ungrasped object position; pick-and-throw adds retention, acceleration, and release along a prescribed direction at or above a specified speed.The finite-horizon tendon-command trajectory is executed open loop at the task level, without object-state or contact-feedback updates.Source: B. Manipulation Task Objectives · p. 2

C. Problem Formulation

  • The method optimizes a tendon-command trajectory against a task-dependent outcome cost using complete physics-simulation rollouts.Robot configuration, object motion, and contact evolution arise from the simulated robot–object–environment interaction.Source: C. Problem Formulation · p. 2

III. METHODOLOGY

The methodology defines task-specific outcome costs, parameterizes tendon commands for search, and reuses prior solutions across task conditions.

III. METHODOLOGY

A. Outcome Costs and Actuation Parameterization

B. Optimize–Learn–Refine Behavior Synthesis

The framework searches finite-dimensional tendon-command parameters for each manipulation condition, then learns a task-conditioned initialization that CMA–ES refines and executes open loop.

B. Optimize–Learn–Refine Behavior Synthesis

  • CMA–ES searches finite-dimensional actuation parameters p through complete rollouts for each condition, retaining the best solution within its search budget.Derivative-free search is used because the contact-rich rollout cost may be nonsmooth.Source: B. Optimize–Learn–Refine Behavior… · p. 4
  • A neural predictor is trained by supervised regression on optimized solutions, which capture how effective actuation varies across task conditions.The learned output is an actuation-parameter initialization, not a solution already evaluated under the new condition.Source: B. Optimize–Learn–Refine Behavior… · p. 4
  • For a new condition, the predicted initialization is reevaluated in simulation, refined by CMA–ES, and used to generate an open-loop tendon-command trajectory.The resulting best-found parameters are condition-specific.Source: B. Optimize–Learn–Refine Behavior… · p. 5

A. Simulation Setup and Data Collection

The simulation study evaluates randomized grasping and directional pick-and-throw in MuJoCo, using separate training and test conditions for the two tasks.

A. Simulation Setup and Data Collection

  • All MuJoCo simulation rollouts used a 3.5 s control horizon for randomized grasping and pick-and-throw.The grasping task used Eq. (7), and pick-and-throw used Eq. (10).Source: A. Simulation Setup and Data Collection · p. 5
  • 84 grasping training positions and 500 test positions were sampled from the same robot-centered workspace, spanning Po,x(0) ∈ [−0.228, 0.227] m and Po,z(0) ∈ [−0.100, 0.450] m.Training positions were paired with CMA–ES-optimized parameters for neural training.Source: A. Simulation Setup and Data Collection · p. 5
  • Pick-and-throw used release directions of 135°, 45°, and 90°, with 10 optimized training conditions and 100 simulation evaluations per setting.The manipulated object was a cylinder with radius R = 15 mm.Source: A. Simulation Setup and Data Collection · p. 5

B. Evaluation Protocol and Compared Methods

The evaluation defines task-specific success criteria and compares neural-warm-started optimizers under matched settings, test conditions, and optimization budgets.

B. Evaluation Protocol and Compared Methods

  • Grasping success required reaching the target and maintaining whole-body enclosure; throwing success required direction error below 5° and release speed at least the prescribed minimum.The minimum speeds were 0.18, 0.32, and 0.24 m s−1 for Trials 1–3, respectively.Source: B. Evaluation Protocol and Compared… · p. 5
  • All optimizer–NN alternatives used the same neural warm start and were evaluated under the same task settings, test conditions, and optimization budget.The compared refiners were CMA–ES, DE, SPSA, TuRBO-1, and Nelder–Mead.Source: B. Evaluation Protocol and Compared… · p. 5
  • Fig. 4 presents simulated task execution sequences, with labeled t denoting normalized execution time.Table II covers baseline comparisons for whole-body grasping and pick-and-throw.Source: Figure 4 · p. 5, Table II · p. 6

C. Simulation Results

CMA–ES + NN performs strongly across simulated grasping and directional throwing, with high success across throwing directions and consistent refinement performance across both tasks.

C. Simulation Results

  • CMA–ES + NN achieved 492/500 successful randomized grasps (98.4%) and throwing success rates of 98%, 97%, and 94% for Trials 1–3.The workspace evaluation is shown in Fig. 5.Source: C. Simulation Results · p. 6, Figure 5 · p. 6
  • Mean throwing direction errors were 2.4°, 2.6°, and 4.4° for Trials 1–3; CMA–ES + NN alone stayed below the 5° success threshold in all three trials.It also had the shortest inference time in each throwing trial.Source: C. Simulation Results · p. 6
  • Across the evaluated NN-initialized pipelines, CMA–ES provided more consistent refinement for both contact-dependent grasping and directional pick-and-throw.The comparisons are summarized in Table II.Source: C. Simulation Results · p. 6, Figure 5 · p. 6

D. Ablation Study

The ablation study tests the enclosure terms and learned initialization in whole-body grasping. The results show that angular coverage, body–object proximity, and learned warm starts each contribute to performance.

D. Ablation Study

  • The ablation isolates enclosure-term and learned-initialization effects on grasping, since enclosure is a prerequisite for pick-and-throw.Evaluating grasping avoids confounding these effects with later release-direction or speed errors.Source: D. Ablation Study · p. 6
  • Removing the angular-sweep term reduced grasp success from 98.4% to 42.8%, the largest degradation among the ablations.Bringing the body close to the object alone was not sufficient; the robot also needed sufficient angular coverage for stable enclosure.Source: D. Ablation Study · p. 6
  • Removing the body–object proximity term reduced grasp success to 62.0%.Angular wrapping without close body–object proximity was likewise insufficient for reliable grasp formation.Source: D. Ablation Study · pp. 6–7
  • Without learned initialization, CMA–ES achieved 78.6% success and required a median 1184 rollouts, versus 816 with initialization.The ablation retained CMA–ES refinement but replaced the task-conditioned initialization with random initialization.Source: D. Ablation Study · pp. 6–7

V. EXPERIMENTAL VALIDATION

The hardware evaluation used SpiRob in a planar workspace, with tendon commands sent open loop and video recordings used for evaluation. Grasping and pick-and-throw success were judged against task-specific criteria, with release velocity estimated from calibrated camera footage.

A. Experimental Setup

  • SpiRob was mounted above a planar workspace and actuated by two displacement-controlled tendon motors.Commands were sent at 100 Hz, with tendon displacement limited to 80 mm s−1; command profiles were temporally rescaled when necessary.Source: A. Experimental Setup · p. 7
  • The hardware evaluation comprised 50 grasping executions and 30 pick-and-throw executions, with 10 repetitions per prescribed throwing direction.An overhead camera recorded interactions for offline evaluation without online feedback.Source: A. Experimental Setup · p. 7

B. Evaluation Metrics

  • Grasping success required capturing and retaining the target; pick-and-throw success also required verifying enclosure, acceleration, and release on video.Throwing success additionally required meeting both the prescribed direction-error and minimum-speed criteria.Source: B. Evaluation Metrics · p. 7
  • Pick-and-throw release speed and direction error were estimated from tracked post-release trajectories in spatially calibrated camera recordings.The reported hardware direction errors and release speeds are averaged over ten executions per prescribed direction.Source: B. Evaluation Metrics · p. 7, Table IV · p. 7

C. Experimental Results

In hardware trials, SpiRob successfully grasped and retained every target, and completed all pick-and-throw executions across the three prescribed directions. The conclusion also reports strong simulation results and identifies closed-loop three-dimensional manipulation as future work.

1) Whole-Body Grasping:

  • 100% of the 50 hardware grasping executions captured and retained the target.Targets varied in planar workspace position; some trials also included a stationary cylindrical object near the target.Source: 1) Whole-Body Grasping · p. 7

2) Pick-and-Throw:

  • 100% of the 30 hardware pick-and-throw executions succeeded, with 10/10 successes for each prescribed direction.Across directions, SpiRob retained the object during acceleration before release.Source: 2) Pick-and-Throw · pp. 7–8

VI. CONCLUSION

  • In simulation, grasping success was 98.4%, while pick-and-throw success was 98%, 97%, and 94% across three directions.The conclusion reports 100% success across 50 hardware grasping and 30 pick-and-throw executions; future work will integrate perception and feedback control for closed-loop three-dimensional tasks.Source: §VI · p. 8

Related papers

Loading 2609.38202v2…