Robotics Today
Transcripción
Robotics Today
Robotics Today Pablo Zegers [email protected] Autonomous Machines Center Facultad de Ingenierı́a y Ciencias Aplicadas Universidad de los Andes Chile 2013 Zegers Robotics Today Robotics Search for an Autonomous Machine Duck of Vaucanson (1739). Zegers Robotics Today What Is Driving Robotics Today? A Roadmap for U.S. Robotics: From Internet to Robotics, March 20, 2013 “Three factors drive the adoption of robots: 1. improved productivity in the increasingly competitive international environment; 2. improved quality of life in the presence of a significantly aging society; and 3. removing first responders and soldiers from the immediate danger/action. Economic growth, quality of life, and safety of our first responders continue to be key drivers for the adoption of robots.” Zegers Robotics Today Important Advances in Mobile Robotics Similar Efficiency to that of Animals I 1 kW at 22km/h. I Three phase permanent magnet synchronous motor that produces twice the torque. MIT Cheetah Robot. Zegers Robotics Today New Products Low Cost Manipulation I Dexterous. I No programming. I Ready to operate. I No cage. I Check www.universal-robots.com too! Baxter, Rethink Robotics. Zegers Robotics Today Hybrid Approaches Leveraging the Best of Everybody! Robots and People Can Work Faster Together, David Bourne, Director Rapid Manufacturing Lab, Robotics Institute, Carnegie Mellon University, July 25, 2013. Zegers Robotics Today Unstructured Object Manipulation A Robotic Frontier I Unseen objects. I Changing geometries. I Difficult to define formally. www.seriouseats.com (June 16, 2011). Zegers Robotics Today Folding Towels (Maitin-Shepard et al, 2010) Pick the Towel Up Maitin-Shepard et al, 2010 I Towel is randomly placed. I Pick it and rotate looking for corners. Zegers Robotics Today Folding Towels (Maitin-Shepard et al, 2010) (cont.) Find Grasp Points Maitin-Shepard et al, 2010 I Look for depth discontinuities consistency through time with the help of a dense sub-pixel optical flow. I Use RANSAC to fit corners to border points and find candidates. I Stereo correspondence for 3D localization. Zegers Robotics Today Folding Towels (Maitin-Shepard et al, 2010) (cont.) Check and Measure Maitin-Shepard et al, 2010 I Pull taut and twist to check grasping. I Fit a rectangle to measure the 3D size. I Start the more structured section of the task: folding. Zegers Robotics Today Folding Towels (Maitin-Shepard et al, 2010) (cont.) Fold Maitin-Shepard et al, 2010 I NVIDIA GTX 295 GPU optimized for dense optical flow. I Intel Core 2 quad core 2.5 GHz CPU. I Average of 1478 seconds per towel, most spent on grasp point detection. Zegers Robotics Today Folding Towels (Maitin-Shepard et al, 2010) (cont.) Finish! Maitin-Shepard et al, 2010 I 100% success rate on 50 previously unseen towels. Zegers Robotics Today Folding Towels (Maitin-Shepard et al, 2010) (cont.) Pipeline Maitin-Shepard et al, 2010 Zegers Robotics Today Folding Towels (Maitin-Shepard et al, 2010) (cont.) Lessons I Sensing and grasping is a single problem. I I I 3D vision is a cornerstone. Grasping is a whole universe. Need of a different hardware. I Unstructured task is composed of structured subproblems. I Autonomous operation but far from autonomous learning. Zegers Robotics Today Grasping Things Typical Task Saxena et al, 2010 I I I Different objects, densities, materials, surfaces, etc. Is it delicate or dangerous? Where should a robot grasp an object? Zegers Robotics Today Grasping Things Grasping Point Detection Saxena et al, 2010 I Many objects are designed to be grasped. I Use synthetic data to train detector. I Use probabilistic model and maximize likelihood in order to infer grasping point position. Zegers Robotics Today Grasping Things Examples Saxena et al, 2010 Zegers Robotics Today Grasping Things Results Saxena et al, 2010 Zegers Robotics Today Grasping Things Lessons I 3D vision and 3D models of reality. I Gripper design. I Lack of real training data because it is a time-consuming task. I Synthetic training. I Some objects require complex sequences in order to be grasped (i.e. book). Zegers Robotics Today Cortando Pan Se Usa Un Instrumento Matı́as Torrealba y Pablo Zegers, Centro de Máquinas Autónomas, Facultad de Ingenierı́a y Ciencias Aplicadas, Universidad de los Andes, Septiembre, 2013 Zegers Robotics Today Cortando Pan El Experimento Matı́as Torrealba y Pablo Zegers, Centro de Máquinas Autónomas, Facultad de Ingenierı́a y Ciencias Aplicadas, Universidad de los Andes, Septiembre, 2013 Zegers Robotics Today Cortando Pan Capturando La Esencia Matı́as Torrealba y Pablo Zegers, Centro de Máquinas Autónomas, Facultad de Ingenierı́a y Ciencias Aplicadas, Universidad de los Andes, Septiembre, 2013 Zegers Robotics Today Cortando Pan Lecciones I Las trayectorias encierran toda la información de la retroalimentación y el control. I Hay caracterı́sticas propias del movimiento humano que no tienen porque imitar un robot. I Hay espacio para hacer las cosas de otra manera. Zegers Robotics Today Conclusion Future Steps I I Model non-linear and non-rigid systems. Integrate sensing and grasping: I I I I Motion primitives: I I I I 3D vision. Grasping technology. Better hardware. Human dynamics offers a good starting point. Determine building blocks. Learn to combine them. Transition from fully human to totally autonomous: I I I I It will take time. Integrate simulators. Annotation tool to generate massive training data from real data. Need an interface to transfer knowledge. Zegers Robotics Today Keywords Tools for Searching I I I I I I I I I I I I I Apprenticeship learning. Demonstration learning. Dynamical primitives. Falling strategies for robots. Grasping. Household tasks. Imitation learning. Learning from others. Movement mapping. Motion primitives. Motor skill learning. Robot learning. Transfer learning. Zegers Robotics Today Who Should Be Watched Not Many! I I I I I I I I I I Pieter Abbeel Sylvain Calinon Auke Ijspeert Jun Morimoto Andrew Ng Jan Peters Torsten Reil Stefan Schaal Peter Stone Russ Tedrake Zegers Robotics Today Bibliography Starting Point! I Abbeel, P., Coates, A., and Ng, A., Autonomous Helicopter Aerobatics through Apprenticeship Learning, The International Journal of Robotics Research, 2010. I Calinon, S., D’Halluin, F., Sauser, E. L., Caldwell, D. G., and Billard, A. G., Learning and Reproduction of Gestures by Imitation: An Approach Based on Hidden Markov Model and Gaussian Mixture Regression, IEEE Robotics & Automation Magazine, June, 2010. I Kober, J., and Peters, J., Imitation and Reinforcement Learning: Practical Algorithms for Motor Primitives in Robotics, IEEE Robotics & Automation Magazine, June, 2010. I Kruger, V., Herzog, D. L., Baby, S., Ude, A., and Kragic, D., Learning Actions from Observations: Primitive-Based Modeling and Grammar, IEEE Robotics & Automation Magazine, June, 2010. I Maitin-Shepard, J., Cusumano-Towner, M., Lei, J., and Abbeel, P., Cloth Grasp Point Detection based on Multiple-View Geometric Cues with Application to Robotic Towel Folding, Proceedings of the International Conference on Robotics and Automation, 2010. I Morimoto, J., Jenkins, O. C., and Toussaint, M., Robot Learning in Practice, IEEE Robotics & Automation Magazine, June, 2010. I Saxena, A., Diemeyer, J., Kearns, J, and Ng, A., Robotic Grasping of Novel Objects, Advances in Neural Information Processing Systems, 2006. I Schaal, S., and Atkeson, C., Learning Control in Robotics: Trajectory-Based Optimal Control Techniques, IEEE Robotics & Automation Magazine, June, 2010. Zegers Robotics Today