Research

Interactive Imitation Learning

Implicitly programming robots

Task Representation Learning

Teaching robots new skills

Planning Alongside Humans

Interactions with human partners

Interactive Imitation Learning

Often times, explicitly programming a robot can be very challenging. Imitation learning offers a more scalable option of implicitly programming robots through demonstrations, interventions or preferences. While we currently have simple algorithms with strong theory, they rely on a set of restrictive assumptions. Notably, an optimal human expert who can interactively provide corrections on any state the robot visits. An everyday human user, however, presents a number of challenges:

1. Mismatched capabilities: A user may demonstrate a task that is beyond the robot’s capabilities. Or, they may not be as efficient as a robot. Can we learn under such mismatch?

2. Unobserved contexts: Human feedback is often influenced by latent context (e.g. trust, attention) that the robot does not directly observe. How do we deal with such uncertainty?

3. Natural feedback: Can we learn from natural modes of human feedback like language and gestures?

Task Representation Learning

We want personal robots that come with a repertoire of skills that can be composed to solve any boutique task in our homes. Every home is different, every human is different. Robots must be able to learn new tasks from a handful of demonstrations and interactions with the human user. Abstractly, one can think of a task as – given an initial configuration of objects, reach a desired goal configuration following a series of feasible operations that do not violate constraints. Learning both the goal and constraints comes with a number of challenges:

1. Skill Composition: Given a library of skills, can we label and learn demonstrated tasks as a composition of known skills?

2. Limited Labels: Getting large amounts of labelled data for every new task can be expensive. How can we leverage unlabelled data collected from multiple tasks that share sub-tasks in common?

3. Learning Structure: Instead of memorizing a task as a sequence of operations, can we learn the underlying structure of the task, i.e. common sub-goals, invariances and dependencies?

Planning Alongside Humans

For robots to work seamlessly alongside human partners, they must operate in a safe and legible manner while matching human cadence. The fundamental challenge is uncertainty. Robots are uncertain about the intent of their human partners, and how this intent changes based on the robot’s actions. Modelling this uncertainty and planning with it in real-time presents a set of challenges:

1. Discrete modes: Theoretically, planning under uncertainty over continuous spaces is intractable. But humans do this everyday by chunking up the continuous space of actions into discrete modes. What are these modes and how do we learn them?

2. Hedging v/s Asserting: In much of driving, humans fluidly trade-off between hedging against uncertain outcomes and assertive actions that collapse uncertainty. How do we learn this trade-off and can we generalize this broadly across human robot interactions?

3. Hierarchy of Abstractions: Robots must deal with uncertainty at multiple time-scales. How can we learn a hierarchy of plannable abstractions to continually build and refine an estimate of the value function?