Graph-Based Inverse Optimal Control for Robot Manipulation

Authors: Arunkumar Byravan, Mathew Monfort, Brian Ziebart, Byron Boots, Dieter Fox

IJCAI 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the effectiveness of the approach with experiments conducted on two 7-degree of freedom robotic arms.
Researcher Affiliation Academia 1Department of Computer Science & Engineering, University of Washington, 2Department of Computer Science, University of Illinois at Chicago, 3School of Interactive Computing, Georgia Institute of Technology
Pseudocode Yes Algorithm 1 explains the procedure for generating the sparse discrete graph.
Open Source Code Yes [Byravan, 2015] Arunkumar Byravan. Project website: Graphbased IOC. http://rse-lab.cs.washington.edu/projects/graphbased-ioc, 2015.
Open Datasets No For each task, we collected demonstrations from multiple users via kinesthetic teaching (Fig. 4a). We record the joint angles and the various object positions (segmented point clouds and bounding boxes) for each of the demonstrations. On average, we collected 50 demonstrations for each task from four different users.
Dataset Splits Yes We split the demonstrations into a training set (70%) for learning the cost function and a held out test set (30%) to measure performance.
Hardware Specification No The paper mentions using 'Barrett WAM and PR2' robots but does not specify any detailed hardware components like GPU or CPU models, or specific cloud computing instances used for training or experiments.
Software Dependencies No The paper discusses various algorithms like CHOMP, STOMP, Traj Opt, and Max Ent IOC, but it does not specify any version numbers for these or other software dependencies.
Experiment Setup Yes We discretize the straight line path into m = 21 points and generate a graph with M = 210000 nodes and k = 15. We initialize the weights (θ) randomly and iterate until convergence. We use two sets of parameters for the Local Trajectory Optimizer. For tasks with constrained motions, we set η = 10 and step-size 1/λ = 0.01 and for larger motions, we set η = 4 and 1/λ = 0.025 for aggressive optimization. We run LTO for 100 iterations. For initialization, we use the least cost discrete path and 25 discrete path samples (interpolated 10x) from the graph, based on the learned cost. For CHOMP, we use a fixed step-size (0.005), 0.1m collision threshold, unit smoothness and obstacle weights and 200 iterations.