Tom Silver

Welcome to my academic website. You are probably here to see pictures of my dog. Otherwise, read on.

I am a final year computer science PhD student at MIT. I am advised by Leslie Kaelbling and Joshua Tenenbaum and am a member of the Learning and Intelligent Systems group and the Computational Cognitive Science group. I am grateful for support from an NSF Graduate Research Fellowship and an MIT Presidential Fellowship. Previously I was a researcher at Vicarious AI. I received my B.A. from Harvard in computer science and mathematics in 2016.

Feel free to contact me at tomssilver@gmail.com.

[Google Scholar] [Twitter]

Job Market

I am on the faculty job market. My CV and statements are below:

[CV] [Research Statement] [Teaching Statement] [Diversity Statement]

News

• (Dec 2023) I am continuing to intern part-time at the Boston Dynamics AI Institute.
• (Dec 2023) "Generalized planning in PDDL domains with pretrained large language models" accepted at AAAI 2024.
• (Aug 2023) "Learning efficient abstract planning models that choose what to predict" accepted at CoRL 2023.
• (Aug 2023) Our Workshop on Learning Effective Abstractions for Planning happened at CoRL 2023.
• (Jul 2023) Our Workshop on Learning for Task and Motion Planning happened at RSS 2023.

Blog Posts

• (Nov 2023) Bilevel Planning for Robots with Nishanth Kumar, Willie McClinton, and Kathryn Le
• (Jul 2023) What's Your Problem? An Oft-Missing Section in AI Papers
• (Mar 2021) Learning to Generate Abstractions for Faster Planning with Rohan Chitnis
• Part 1: Learning to Generate Context-Specific Abstractions for Factored MDPs
• Part 2: Learning to Generate Abstractions in Problems with Relational Structure
• (Apr 2018) Lessons from My First Two Years of AI Research

Research

My research is motivated by the prospect of broadly-competent intelligent robots that can respond to very high-level commands like "make me a heart-healthy dinner"; learn new skills like "grind fresh pepper"; and learn new concepts like "wilted spinach." Such robots would be especially transformative for people who could not otherwise remain independent in their homes. Most of my work is at the intersection of automated planning and machine learning: learning to plan and planning to learn while making efficient use of limited data and time. I often use techniques from task and motion planning, program synthesis, and neuro-symbolic machine learning. Below are selected works; see Google Scholar for a full list.

Learning Abstractions for Robot Planning

Abstractions can be very useful for decision making because they allow the agent to first focus on the high-level aspects of a task before getting bogged down in details. We would like a robot to synthesize abstractions — state abstractions (predicates), action abstractions (skills), and task abstractions (relevant objects) — that are most useful for planning in its specific domain.

Learning State and Action Abstractions

We learn neuro-symbolic and relational state and action abstractions from demonstrations. The abstractions are explicitly optimized for effective and efficient bilevel planning.

We learn neuro-symbolic skills with goal-conditioned policies from demonstrations and symbolic predicates. The learned skills can be used with bilevel task and motion planning techniques.

We study generalized planning in robotic settings that involve both discrete and continuous actions. We propose a method for learning state and action abstractions that leads to much faster decision-making over planning from scratch.

We propose a bottom-up relational approach for learning operators for task and motion planning. Our approach can be seen as a "model-based" learning approach to TAMP. We compare with several model-free baselines.

Learning Task Abstractions

We propose a graph neural network architecture for predicting object importance to address the challenge of planning in problems that contain many, many objects.

We observe that learning to impose constraints in factored planning problems can induce context-specific abstractions that afford much more efficient planning.

Learning Policies for Rapid Decision Making

Planning is important when the robot is faced with difficult and unfamiliar problems. But once the robot gets into the habit of solving the same kinds of problems repeatedly, its decision making should become fast and "second nature." In other words, the robot should compile its planning experience into a reactive policy.

We use GPT-4 to synthesize Python generalized plans for PDDL domains. We also propose an automated debugging scheme that dramatically improves performance over one-time prompting.

We propose a new method for generalized policy search. The main idea is that a candidate policy should be used to guide planning on training problems as a mechanism for evaluating that candidate.

We can learn policies from five or fewer demonstrations that generalize to dramatically different test task instances.

We present a simple method for improving nondifferentiable policies using model-free deep reinforcement learning.

Learning to Improve Decision Making through Online Exploration

An intelligent robot should actively collect its own data, and learn from that data, so that it can make better decisions in the future. If an expert is available to provide demonstrations or answer questions, the robot should ask for help, but should do so judiciously.

We consider a setting where a robotic agent must learn the meaning of predicates by querying an expert about the current world state. Before generating an informative query, the robot must take actions to reach an interesting world state.

We propose goal-literal babbling (GLIB), a simple and general method that addresses the problem of efficient exploration for transition model learning in the relational model-based reinforcement learning setting without extrinsic goals or rewards.

We present PDDLGym, a framework that automatically constructs OpenAI Gym environments from PDDL domains and problems.

We introduce the Schema Network, an object-oriented generative physics simulator capable of disentangling multiple causes of events and reasoning backward through causes to achieve goals.

Template from here.