Siyuan, Feng PhD Student The Robotics Institute Carnegie Mellon University 5000 Forbes Avenue Pittsburgh, PA 15213 (412) 482-8651 sfeng [at] cs.cmu.edu

Here is my cv.

Research:

I am generally interested in using optimization and learning techniques to generate complex robot behaviors. My current research focuses on developing rough terrain locomotion and full body manipulation capabilities for humanoid robots using online optimization.

For a useful general purpose humanoid robot, I think it is important for the user or a higher level planner to be able to specify a set of desired foot steps. This capability is especially useful in a cluttered environment where obstacle avoidance is necessary. Thus, it is ideal for the robot controller to adapt to the given foot steps during runtime. I am taking an hierarchical approach to this problem:

• The high level controller approximates the robot using a simple point mass model, and optimizes a center of mass trajectory based on the given desired foot steps. This controller typically runs per foot step.
• The mid level controller modifies the high level trajectory using Receding Horizon Control in case of large disturbances. Compared with the high level controller, it optimizes a much shorter trajectory, but enforces constraints and reoptimizes on every time step.
• The low level controller runs on every time step, and uses convex optimization to generate joint space commands that best track the modified trajectory while obeying full robot dynamics and physical constraints on each time step.

Dynamic walking on Atlas:

Here are some examples of Atlas walking under strong perturbation using the proposed approach. The mid level controller is optimizing for the swing foot placement that best tracks the planned center of mass trajectory.

Full body manipulation on Atlas:

One advantage of humanoid robots is having large degrees of freedom (DoF) with a relatively small footprint, which can be beneficial for manipulation tasks in tight spaces. On the other hand, having large DoF requires the controller to resolve redundancy intelligently and makes motion planning harder in general. Maintaining balance also becomes challenging. For manipulation, the same low level controller for walking is used. It greedily tracks desired joint space or Cartesian space accelerations for specific body parts while maintaining balance. To perform more complex tasks, a separate motion planner is used to generate full body trajectories, which are then executed by the controller.

The DARPA Robotics Challenge:

In the DARPA Robotics Challenge (DRC), individual teams develop robot systems to perform disaster response-like tasks. I participated as the control lead of Team WPI-CMU. For the competition, we used the Atlas robot, a full-size hydraulic humanoid robot developed by Boston Dynamics. I have developed full body controllers for Atlas to perform locomotion and manipulation tasks.

At the DRC Finals, each team had two chances, and the best run is used for ranking. Team WPI-CMU finished with 7 out of 8 points for both runs. The following are sped up videos of our second run at the Finals. We were unable to get the drill point due to a hardware failure in the right forearm, which the robot used for drilling. It overheated due to an faulty encoder when the robot was lifting its right arm. The uncontrolled right arm wedged the drill into the wall and generated a large backward force on the robot. Our fall detection algorithm promptly kicked in and prevented the robot from falling. The operator then manually recovered and moved on. We lost more than six minutes right in front of the stairs due to an unintended communication blackout on the DARPA side. Communication was eventually restored, but the lost time was not honored.

Nevertheless, we were the only humanoid team that had no falls or physical human interventions (resets) during the two runs at the Finals. We also had the most consistent performance out of all the competitive teams.

Other videos of the Atlas robot:

Publications:

1. S. Feng, X Xinjilefu, C. Atkeson and J. Kim. Optimization Based Controller Design and Implementation for the Atlas Robot in the DARPA Robotics Challenge Finals - Video. Proceedings of IEEE-RAS International Conference on Humanoid Robots, Seoul, Korea, 2015.


2. X. Xinjilefu, S. Feng and C. Atkeson. Center of Mass Estimator for Humanoids and its Applica- tion in Modelling Error Compensation, Fall Detection and Prevention. Proceedings of IEEE-RAS International Conference on Humanoid Robots, Seoul, Korea, 2015.


3. C. Liu, C. Atkeson, S. Feng and X. Xinjilefu. Full-body Motion Planning and Control for The Car Egress Task of the DARPA Robotics Challenge. Proceedings of IEEE-RAS International Conference on Humanoid Robots, Seoul, Korea, 2015.


4. S. Feng, E. Whitman, X Xinjilefu, and C. Atkeson. Optimization-based Full Body Control for the DARPA Robotics Challenge. Journal of Field Robotics, Volume 32, Issue 2, pages 293-312, March 2015.


5. M. DeDonato, V. Dimitrov, R. Du, R. Giovacchini, K. Knoedler, X. Long, F. Polido, M. A. Gennert, T. Padir, S. Feng, H. Moriguchi, E. Whitman, X Xinjilefu, and C. Atkeson. Human-in-the-loop Control of a Humanoid Robot for Disaster Response: A Report from the DARPA Robotics Challenge Trials. Journal of Field Robotics, Volume 32, Issue 2, pages 275-292, March 2015.


6. S. Feng, E. Whitman, X Xinjilefu, and C. Atkeson. Optimization Based Full Body Control for the Atlas Robot - Video. In Proceedings of IEEE-RAS International Conference on Humanoid Robots, Madrid, Spain, 2014.


7. X Xinjilefu, S. Feng, and C. Atkeson. Dynamic State Estimation Using Quadratic Programming. In Proceedings of IEEE/RSJ International Conference on Intelligent Robots and Systems, Chicago, IL, 2014.


8. X Xinjilefu, S. Feng, W. Huang, and C. Atkeson. Decoupled State Estimation for Humanoids Using Full-body Dynamics. In Proceedings of IEEE International Conference on Robotics and Automation, Hong Kong, China, 2014.


9. S. Feng, X Xinjilefu, W. Huang, and C. Atkeson. 3D Walking Based on Online Optimization (Best Oral Paper Award) - Video. In Proceedings of IEEE-RAS International Conference on Humanoid Robots, Atlanta, GA, 2013.


10. A. Degani, A. Long, S. Feng, H. B. Brown, R. D. Gregg, H. Choset, M. T. Mason, and K. Lynch. Design and Open-Loop Control of the ParkourBot, a Dynamic Climbing Robot. IEEE Transactions on Robotics, 30(3):705-718, 2014.


11. A. Degani, S. Feng, H. B. Brown, K. Lynch, H. Choset, and M. T. Mason. The ParkourBot - A Dynamic BowLeg Climbing Robot. In Proceedings of IEEE International Conference on Robotics and Automation, Shanghai, China, 2011.


12. A. Degani, S. Feng, H. Choset, and M. T. Mason. Minimalistic, Dynamic, Tube Climbing Robot - Video (Best Video Award) . In Proceedings of IEEE International Conference on Robotics and Automation, Anchorage, AK, 2010.