Learning and leveraging kinematics for robot motion planning under uncertainty




Jain, Ajinkya

Journal Title

Journal ISSN

Volume Title



Service robots that can assist humans in performing day-to-day tasks will need to be general-purpose robots that can perform a wide array of tasks without much supervision from end-users. As they will be operating in unstructured and ever-changing human environments, they will need to be capable of adapting to their work environments quickly and learning to perform novel tasks within a few trials. However, current robots fall short of these requirements as they are generally highly specialized, can only perform fixed, predefined tasks reliably, and need to operate in controlled environments. One of the main reasons behind this big gap is that the current robots require complete and accurate information about their surroundings to function effectively, whereas, in human environments, robots will only have access to limited information about their tasks and environments. With incomplete information about its surroundings, a robot using pre-programmed or pre-learned motion policies will fail to adapt to the novel situations encountered during operation and fall short in completing its tasks. Online motion generation methods that do not reason about the lack of information will not suffice either, as the developed policies may be unreliable under incomplete information. Reasoning about the lack of information becomes critical for manipulation tasks a service robot would have to perform. These tasks will often require interacting with multiple objects that make or break contacts during the task. A contact between objects can significantly alter their subsequent motion and lead to sudden transitions in their dynamics. Under these sudden transitions, even minor errors in estimating object poses can cause drastic deviations from the robot's initial motion plan for the task and lead the robot to failure in completing the tasks. Hence, service robots need methods that generate motion policies for manipulation tasks efficiently while accounting for the uncertainty due to incomplete or partial information.

Partially Observable Markov Decision Processes (POMDPs) is one such mathematical framework that can model and plan for tasks where the agent lacks complete information about the task. However, POMDPs incur exponentially increasing computational costs with planning time horizon, which restricts the current POMDP-based planning methods to problems having short time horizons. Another challenge for planning-based approaches is that they require a state transition function for the world they are operating in to develop motion plans, which may not always be available to the robot. In control theory terms, a state transition function for the world is analogous to its system plant. In this dissertation, we propose to address these challenges by developing methods that can learn state transition functions for robot manipulation tasks directly from observations and later use them to generate long-horizon motion plans to complete the task under uncertainty.

We first model the world state transition functions for robot manipulation tasks involving sudden transitions, such as due to contacts, using hybrid models and develop a novel hierarchical POMDP-planner that leverages the representational power of hybrid models to develop motion plans for long-horizon tasks under uncertainty. Next, we address the requirement of planning-based methods to have access to world state transition functions. We introduce three novel methods for learning kinematic models for articulated objects directly from observations and present an algorithm to construct the state transition functions from the learned kinematics models for manipulating these objects. We focus on learning models for articulated objects as they form one of the biggest sets of household objects that service robots will frequently interact with. The first method, MICAH, focuses on learning kinematic models for articulated objects that exhibit configuration-dependent articulation properties, such as a refrigerator door that stays closed magnetically, from unsegmented sequences of observations of object part poses. Next, we introduce ScrewNet, which removes the requirement of object pose estimation of MICAH and learns articulation properties of objects directly from raw sensory data available to the robot (depth images) without knowing their articulation model category a priori. Extending it further, we introduce DUST-net, which learns distributions over articulation model parameters for objects indicating the network's confidence over the estimated parameters directly from raw depth images. Combining these methods, in this dissertation, we introduce a unified framework that can enable a robot to learn state transition functions for manipulation tasks from observations and later use them to develop long-horizon plans even under uncertainty.


LCSH Subject Headings