Research

Mobile-Robot Error and Recovery

When a mobile robot gets stuck, the robot's conventional locomotion strategy and the structured dynamics originally used to model the motion are no longer applicable. How can the robot "escape"? To solve this problem, we first identify the dominant elements of the "new" dynamics, and then propose unconventional use of existing mechanisms to produce locomotion. We also provide structure to the mobile-robot error domain by classifiying locomotion errors based on the presence of undesired contact or the absence of a desired contact. For example, a legged robot requires sufficient traction between it's feet and the ground. If the ground is slippery, the robot does not have the desired contact between the feet and the ground, and the robot is stuck. If the robot is high-centered, then the robot has an undesired contact between the ground and the body, and the robot is stuck. Thus, depending on the situation, we require a locomotion strategy that returns the robot to its conventional contact mode. This research germninated during my summer 2002 internship in the University of Michigan with Prof. Daniel Koditschek. Here are some interesting movies of RHex getting stuck:

1) Crossing a region littered with deep holes (successful)
2) Negotiating an array of tall rocks (difficult, gets high-centered)
3) Escaping when the body is stuck between rocks (successful)
4) Climbing over a cinderblock wall (successful when using dynamic maneuver)
5) Crossing a region littered with compliant crevices (successful)
6) Crossing a region comprising compliant and rigid obstacles (difficult, requires quasistatic and dynamic maneuvers)
7) Escaping when body is stuck between compliant obstacles (successful)

Legless Locomotion

A flipped turtle is stuck because it's feet do not touch the ground. One way of escape is to rock itself on it's shell so that it rights itself. Alternately, can the turtle locomote in it's flipped configuration merely by swinging the legs and rocking back and forth? Similarly, what happens when a robot is high-centered? How can it move in such a situation? My doctoral work has "discovered" this novel locomotion technique for high-centered robots, a locomotion we call legless locomotion.

To simplify the problem, we built a prototype called Rocking and Rolling Robot (RRRobot). It is "high-centered for life", that is, its legs don't touch the ground in its nominal position. We show in a recent paper that it is possible to locomote RRRobot just by wiggling its short legs, i.e., the robot locomotes by transferring energy and momentum to produce a rocking and rolling body motion. This rocking motion when coupled with the contact constraints produces incremental translation. This work was a finalist in the Best Student Paper competition at ICRA, 2004 .

Check out this movie that summarizes the project. The following figure showing RRRobot's planar motion captured using a Vicon system in experiment (the arrows indicate robot orientation). A big thanks to CMU's mocap lab. Here is an animation of the same motion (10X speed; it plays well in Windows Media Player in Windows).

An important aspect of this locomotion is that RRRobot's inertia varies with leg configuration. In a forthcoming paper, we will provide insights to an inverse-dynamics solution, a step closer to the first planning technique for a variable-inertia robot that locomotes using shape changes in the presence of nonholonomic constraints.

Automatic Gait Generation using Dynamic Programming

Suppose a car with an active suspension is stuck in a slippery ditch. The goal is to get the car out of the ditch. This problem is a slight variation of the famous car-hill AI problem, but with a larger state and input space. Here, the car mass can move on its suspension. It turns out that the problem dynamics is quite complex, partly due to the coriolis forces that arise due to sprung mass velocity and wheel velocity. Can we induce the car to escape without slipping? In addition to a bare-bones Newton's law analysis, we use dynamic programming (Konkimalla and Lavalle, IJRR, 2001) to search for a solution. The cost function that was optimized was a weighted sum of car mass deviation and inputs.

Kinematic Reduction of Dynamic Systems

Planning and gait synthesis for dynamic systems is difficult because of the presence of drift terms in the equations and the controls are forces or accelerations. In contrast, planning and gait synthesis are much simpler for kinematic systems. Using prior work by Bullo, Lewis, and Lynch, we explore kinematic reductions for legless locomotion.

Anthropomorphic Manipulation

I interned at Anybots, Inc. in summer 2004, where my work involved developing teleoperation for controlling a seven degree-of-freedom arm and a twenty degree-of-freedom hand. I built GUIs in Python and researched existing MoCap solutions. We ultimately chose a magnetic MoCap product and I developed inverse kinematics solutions for real-time position control.

Previous work

I participated in RoboCup 2001 as part of the CMU-Dragons Small Size Robots team. I designed, built, and maintained a set of omni-directional robots. Some of the salient features of this robot design is that it uses special mechanum wheels and is extremely light and fast. I also contributed various defence and attack soccer plays for the heterogeneous robot team.

In 2001-02, I built a planar version of the Goes-Over-All-Terrain (GOAT) robot and a quasistatic planner to study locomotion and manipulation capabilites of a fully-actuated robot. The GOAT has four legs, with wheels at the end of each leg. All legs and wheels are directly actuated. See Tucker Balch's goat page for more details.

Selected Publications

1. Ravi Balasubramanian, Alfred A. Rizzi, and Matthew T. Mason. A Dynamic Feedback Strategy Using Active Suspension for escaping from a rut. (In preparation).
2. Ravi Balasubramanian, Alfred A. Rizzi, and Matthew T. Mason. Kinematic Reduction and Planning using Symmetry for a Variable Inertia Mechanical System. In Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems, 2004. PDF
3. Ravi Balasubramanian, Alfred A. Rizzi, and Matthew T. Mason. Legless Locomotion: Models and Experimental Demonstration. The IEEE/RSJ International Conference on Robotics and Automation, 2004 (Best Student Paper Finalist). PDF
4. Ravi Balasubramanian, Brendan R. Meeder, Alfred A. Rizzi, and Matthew T.Mason. Legless Locomotion for Legged Robots (MPEG Video). The IEEE/RSJ International Conference on Robotics and Automation, 2004. MPEG
5. Ravi Balasubramanian. Legless Locomotion: Concept and Analysis. 2004. PDF
6. R. Balasubramanian and T. Balch. Energy-Optimal Trajectories for Overactuated Robots. Tech. report CMU-RI-TR-02-17, Robotics Institute, Carnegie Mellon University, July, 2002.