Our friend Sumeet Singh
recently returned from the Robotics (RSS) conference in Pittsburgh, so we asked him to share with us a round-up of the research he’s most excited by.
Sumeet and his colleagues presented their latest research on Risk-Sensitive Inverse Reinforcement Learning (IRL) at the RSS Pioneers Workshop
. Generally, human-robot interactions assume that any decision maker exhibits some degree of risk-aversion to curtail the effect of “bad” outcomes due to uncertainty. However, in safety-critical systems, decision makers usually exhibit some measure of risk-aversion to avoid “bad” outcomes that may be caused by uncertainty in the behavior of other agents. Their paper
explores a set of risk-sensitive Inverse Reinforcement Learning algorithms that can infer, predict, and mimic risk-sensitive behavior, which is crucial for safety-critical applications where humans and robots interact.
Here are 3 of Sumeet’s favorite papers from RSS:
The objective of this paper is to analyze and exploit the immense combinatorics of sequential manipulation and tool-use in complex domains. Imagine trying to pull towards you a ball sitting on the opposite end of a long table using tools such as a couple of hooks. A feasible solution could be to use the hook to slide the ball toward you. Alternatively, one could use the hook to hit the ball towards you and catch it with your other hand. This multimodality is incorporated within the paper’s solution approach by formulating the problem as joint task and motion planning, and using logic predicates and stable maneuver primitives to model the “skeleton” or mode sequence of the solutions. The algorithm then jointly optimizes over this discrete mode sequence and the continuous configuration variables, resulting in quite unexpected and amusing manipulations.
This paper presents a motion planning framework for in-hand manipulation of objects within the gravity plane. While the horizontal planar pusher manipulation task has been well studied, this work extends the notion of “motion cones” (i.e. set of feasible motions for the manipulated object resulting from a frictional pusher and a support surface) to the vertical plane. Using elegant approximations of these “motion cones,” the work presents an efficient sampling-based planner that chains together these motion primitives to achieve impressive in-hand manipulations that additionally leverage interaction with the environment.
Motion planning in known environments has been exhaustively studied using a variety of techniques. However, the class of motion planning problems in unknown or partially known environments remains a substantial challenge that has been less explored (pun intended!). While there are several heuristic methods proposed in the literature, this paper strives to formally establish the theoretical underpinnings for this class of planning problems (e.g., notions of optimality). Building upon the constraint of safe planning to ensure that a robot will never collide even as it explores more of its environment, the paper presents a class of pseudo-optimal policies that automatically balance exploitation (cost optimality) versus exploration (environment discovery).