GOMP: Grasp-Optimized Motion Planning for Bin Picking

Ichnowski, Jeffrey; Danielczuk, Michael; Xu, Jingyi; Satish, Vishal; Goldberg, Ken

doi:10.1109/icra40945.2020.9197548

Cited by 38 publications

(14 citation statements)

References 64 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…2, we compare the shifted geometric mean of solving 10 problems of 20 different dimension, for a total of 200 runs per class per solver. The problem dimensions for Control, Huber, SVM, Lasso are (10,11,12,13,14,16,17,20,23,26,31,37,45,55,68,84,105,132,166,209); for Random and Eq are (10,11,12,13,15,18,23,29,39,53,73,103,146,211,304 Figure 3: Solve time with increasing dimension on the Random QP problem set. We train and benchmark two vector RL adaptation policies: (dashed) on problems ranging from dimension 10 to 50, and (solid) on problems ranging from 10 to 2000.…”

Section: Multi-task/general Rlqp Policymentioning

confidence: 99%

“…Many applications in control [26] and optimization [27] require QPs from the same class to be repeatedly solved. To test if training a policy specific to a QP class can outperform a policy trained on the benchmark suite, we train policies specific to the problems generated by the trust-region [8] based solver for sequential quadratic program (SQP) from a grasp-optimized motion planner (GOMP) [26,25] for robots. With these problems, RLQP trained on the benchmarks converges more slowly than the handcrafted policy included in OSQP.…”

Section: Training a Class-specific Policymentioning

confidence: 99%

See 1 more Smart Citation

Accelerating Quadratic Optimization with Reinforcement Learning

Ichnowski

Jain

Stellato

et al. 2021

Preprint

Self Cite

View full text Add to dashboard Cite

First-order methods for quadratic optimization such as OSQP are widely used for largescale machine learning and embedded optimal control, where many related problems must be rapidly solved. These methods face two persistent challenges: manual hyperparameter tuning and convergence time to high-accuracy solutions. To address these, we explore how Reinforcement Learning (RL) can learn a policy to tune parameters to accelerate convergence. In experiments with well-known QP benchmarks we find that our RL policy, RLQP, significantly outperforms state-of-the-art QP solvers by up to 3x. RLQP generalizes surprisingly well to previously unseen problems with varying dimension and structure from different applications, including the QPLIB, Netlib LP and Maros-Mészáros problems. Code for RLQP is available at https://github.com/berkeleyautomation/rlqp.

show abstract

Section: Multi-task/general Rlqp Policymentioning

confidence: 99%

Section: Training a Class-specific Policymentioning

confidence: 99%

Accelerating Quadratic Optimization with Reinforcement Learning

Ichnowski

Jain

Stellato

et al. 2021

Preprint

Self Cite

View full text Add to dashboard Cite

show abstract

“…In this section, we describe a trajectory time-optimization to improve the speed of task performance. In prior formulations 60,61 of pick-and-place trajectory optimization, a timeminimized trajectory is found by discretizing the trajectory into a sequence of waypoints and formulating a sequential quadratic program that minimizes the sum-of-squaredacceleration or sum-of-squared distance between the waypoints. We observe that this prior formulation, while versatile enough for the peg transfer tasks, can be simplified to minimizing a sequence of splines defined by up to four waypoints, and relying on the kinematic design of the robot to avoid collisions.…”

Section: Non-linear Trajectory Optimizationmentioning

confidence: 99%

Automating Surgical Peg Transfer: Calibration with Deep Learning Can Exceed Speed, Accuracy, and Consistency of Humans

Hwang,

Ichnowski,

Thananjeyan

et al. 2020

Preprint

Self Cite

View full text Add to dashboard Cite

We consider the automation of the well-known peg-transfer task from the Fundamentals of Laparoscopic Surgery (FLS). While human surgeons teleoperate robots to perform this task with great dexterity, it remains challenging to automate. We present an approach that leverages emerging innovations in depth sensing, deep learning, and Peiper's method for computing inverse kinematics with time-minimized joint motion. We use the da Vinci Research Kit (dVRK) surgical robot with a Zivid depth sensor, and automate three variants of the peg-transfer task: unilateral, bilateral without handovers, and bilateral with handovers. We use 3D-printed fiducial markers with depth sensing and a deep recurrent neural network to improve the precision of the dVRK to less than 1 mm. We report experimental results for 1800 block transfer trials. Results suggest that the fully automated system can outperform an experienced human surgical resident, who performs far better than untrained humans, in terms of both speed and success rate. For the most difficult variant of peg transfer (with handovers) we compare the performance of the surgical resident with performance of the automated system over 120 trials for each. The experienced surgical resident achieves success rate 93.2 % with mean transfer time of 8.6 seconds. The automated system achieves success rate 94.1 % with mean transfer time of 8.1 seconds. To our knowledge this is the first fully automated system to achieve "superhuman" performance in both speed and success on peg transfer. Supplementary material is available at https://sites.google.com/view/surgicalpegtransfer.

show abstract

“…The objective of the QP minimizes jerk. To make the trajectory as fast as possible, in a manner similar to GOMP [10], we repeatedly reduce H by 1 until the QP is infeasible, and use the shortest feasible trajectory. The QP takes the following form:…”

Section: Min Jerk Trajectory Generation Functionmentioning

confidence: 99%

Robots of the Lost Arc: Self-Supervised Learning to Dynamically Manipulate Fixed-Endpoint Cables

Zhang¹,

Ichnowski²,

Seita³

et al. 2020

Preprint

Self Cite

View full text Add to dashboard Cite

High-speed arm motions can dynamically manipulate ropes and cables to vault over obstacles, knock objects from pedestals, and weave between obstacles. In this paper, we propose a self-supervised learning pipeline that enables a UR5 robot to perform these three tasks. The pipeline trains a deep convolutional neural network that takes as input an image of the scene with object and target. It computes a 3D apex point for the robot arm, which, together with a task-specific trajectory function, defines an arcing motion for a manipulator arm to dynamically manipulate the cable to perform a task with varying obstacle and target locations. The trajectory function computes high-speed minimum-jerk arcing motions that are constrained to remain within joint limits and to travel through the 3D apex point by repeatedly solving quadratic programs for shorter time horizons to find the shortest and fastest feasible motion. We experiment with the proposed pipeline on 5 physical cables with different thickness and mass and compare performance with two baselines in which a human chooses the apex point. Results suggest that the robot using the learned apex point can achieve success rates of 81.7% in vaulting, 65.0% in knocking, and 60.0% in weaving, while a baseline with a fixed apex across the three tasks achieves respective success rates of 51.7%, 36.7%, and 15.0%, and a baseline with human-specified task-specific apex points achieves 66.7%, 56.7%, and 15.0% success rate respectively. Code, data, and supplementary materials are available at https: //sites.google.com/berkeley.edu/dynrope/home. I. INTRODUCTIONDynamic manipulation and management of linear deformable objects such as ropes, chains, vacuum cords, charger cables, power cables, tethers, and dog leashes are common in daily life. For example, a person vacuuming may find that the vacuum's power cord is stuck on a chair, and could use dynamic manipulation to "vault" the cord over the chair. If the first motion does not succeed, the human can try again, adapting their motion to the failure. In this paper, we explore how a robot can teach itself to perform three tasks: Task 1: Vaulting The robot dynamically manipulates a cable to move from one side of an obstacle to another. Task 2: Knocking The robot dynamically manipulates the cable to knock a target object off an obstacle. Task 3: Weaving The robot dynamically manipulates the cable to weave it between three obstacles.Hereafter, we use the term cable to refer to any 1D deformable object with minimal stiffness, such as ropes, cords, and strings. To reduce the complexity of parameterizing actions, we compute trajectories using DJ-GOMP [10], [9] that takes as input the apex point in the trajectory. This point,

show abstract

GOMP: Grasp-Optimized Motion Planning for Bin Picking

Cited by 38 publications

References 64 publications

Accelerating Quadratic Optimization with Reinforcement Learning

Accelerating Quadratic Optimization with Reinforcement Learning

Automating Surgical Peg Transfer: Calibration with Deep Learning Can Exceed Speed, Accuracy, and Consistency of Humans

Robots of the Lost Arc: Self-Supervised Learning to Dynamically Manipulate Fixed-Endpoint Cables

Contact Info

Product

Resources

About