I'm thrilled to join Princeton's faculty as an assistant professor in the ECE department starting Fall 2025 🐯
Stay tuned for the launch of my lab. We will develop generally helpful robots that learn and plan 🤖
I defended my PhD
@MITEECS
this week! Thanks to everyone who came out. And thanks especially to
@nishanthkumar23
who not only managed the Zoom, but also got me this amazing gift…
New preprint: "Learning Neuro-Symbolic Skills for Bilevel Planning" w/ Ashay Athalye, Josh Tenenbaum, Tomas Lozano-Perez and Leslie Kaelbling.
Check it out if you're interested in hierarchical RL, TAMP, and LfD!
Paper:
Video:
Our paper got an oral at CoRL (rhyme!) We proposed CAMPs, a method for learning to generate abstractions of large planning problems. Stop by at
@corl_conf
this week!
Paper:
Video:
Code:
Looking forward to the Foundation Models for Decision Making Workshop at
#NeurIPS2022
!
We'll be presenting preliminary work on using large language models for PDDL planning:
These papers also look very interesting & related:
Looking forward to
#AAAI23
! On Tuesday, I'll present work on neuro-symbolic learning for robotic planning at the Bridge Session on AI & Robotics.
I'll show some clips from this 1972 video of Shakey the robot and ask: how much progress have we really made?
New work: "Few-Shot Bayesian Imitation Learning with Logic over Programs"
w/ Kelsey Allen, Alex Lew, Leslie Kaelbling, Josh Tenenbaum
Website:
Short versions to appear at ICLR SPiRL workshop and RLDM (say hello!)
If you're interested in LLMs for planning, check out our new preprint: "Generalized Planning in PDDL Domains with Pretrained Large Language Models"
w/ Soham Dan, Kavitha Srinivas, Josh Tenenbaum, Leslie Kaelbling, Michael Katz (
@MITIBMLab
)
I'm really excited about this "planning to practice" direction. We're able to leave Spot alone for hours and it's much improved when we get back. This is a step towards generalist robots that learn to specialize *during deployment*. Check out the paper!
Can we get robots to improve at long-horizon tasks without supervision?
Our latest work tackles this problem by planning to practice!
Here's a teaser showing initial task -> autonomous practice -> eval (+ interference by a gremlin👿)
I'm trying to collect recent benchmarks for generalization in RL. Below are the ones that I've found so far. Please reply if you know of others, or if I'm missing a state of the art.
This year, Rohan Chitnis and I have been trying to figure out how we can learn models through environment interaction, like in model-based RL, and then use the models to plan w/ powerful robotic planners, like those found in task and motion planning (TAMP).
We are organizing the RSS’23 Workshop on Learning for Task and Motion Planning
Contributions of short papers or Blue Sky papers are due May 19th, 2023.
When is it okay to use a 'max' in reporting results for RL? Two examples:
1. Train on multiple random seeds and report only the best seed
2. At training iteration t, report the max over training iterations 1, 2, ..., t.
(1/10)
(1/2) New work on "Residual Policy Learning" with Kelsey Allen, Josh Tenenbuam & Leslie Kaelbling:
Simple idea: start with a meh policy, learn a residual function to improve it. Does better than deep RL or initial policy alone.
The submission portal for our RSS workshop on learning for TAMP is now open! The deadline is May 19. We're going to have some really exciting speakers. The format is hybrid, so even if you can't make it to Korea, consider submitting!
I'm really excited for our upcoming
#CoRL2023
workshop on learning effective abstractions for planning (LEAP)! The deadline to submit a paper is coming up quickly (Sep 30). We're accepting submissions in both the CoRL or ICRA formats. Hope to see you there!
Our previous work made the crucial assumption that relational state abstractions, in the form of logical predicates, were provided. The main contribution in our new preprint is to learn these predicates from data, without any supervision on their form or number.
What paper do you come back to most often for another read?
For me, it is "From Skills to Symbols: Learning Symbolic Representations for Abstract High-Level Planning" by George Konidaris, Leslie Kaelbing, and Tomas Lozano-Perez. ()
New blog posts (two for the price of one!)
Rohan Chitnis and
@tomssilver
discuss their recent work on learning to generate abstractions for faster planning.
Excited to share this new blog post led by
@nishanthkumar23
@williebeit
and Kathryn Le! It's a great primer on bilevel planning, the bedrock of our learning and planning work over the last few years.
Ever heard about "Bilevel Planning" or "Task and Motion Planning", but been unsure what those words mean? Ever wanted a gentle intro to these methods so you can just understand what's going on? Our new blog post might help!
“Hi Tom -
This is a bit of a weird one. It turns out that in Ubuntu 22.04 there's now a system user, related to the Trusted Platform Module, that is also called 'tss', and that's breaking your ability to log in.”
Time to legally change my name…
Next (), we considered neuro-symbolic techniques for learning two additional components: (1) samplers, which stochastically refine the abstract actions into low-level controllers; and (2) a low-level transition model.
We finally came up with an objective for learning predicates that is efficient enough to evaluate, but still closely tied to our real planning objective. The main idea is to use demonstration data to cheaply approximate the most expensive parts of bilevel planning.
This is joint work with
@nishanthkumar23
,
@williebeit
, Tomas Lozano-Perez, Leslie Kaelbling, and Josh Tenenbaum. Also, we’re looking forward to presenting an extended abstract at
#RLDM2022
!
Predicate learning proved to be the most challenging part of the pipeline. Among many issues that we encountered, one interesting theme was: state abstractions that are good for making predictions (bisimulation) are not necessarily good for planning.
Hello, world! We are the Learning and Intelligent Systems group
@MIT_CSAIL
, headed by Leslie Pack Kaelbling & Tomás Lozano-Pérez. We work on AI, ML, and robotics, and we’ll be mostly tweeting about new work by our group.
Our objective has been to learn all of the models needed for bilevel planning, as in task and motion planning (TAMP). Unlike in model-based reinforcement learning, a low-level transition model is not enough; we also need state and action abstractions for high-level planning.
Then on Saturday, I'll present our work on "Predicate Invention for Bilevel Planning" at the main conference (oral at 9:30am in Room 147B; poster at 6:15pm).
Let me know via email if you're at
#AAAI23
and want to chat!
There was a lot of good and interesting debate on "is scaling all we need to solve robotics?" at
#CoRL23
. I spent some time writing up a blog post about all the points I heard on both sides:
Also check out
@nishanthkumar23
and
@williebeit
's
#CoRL2023
paper, the latest advance in our effort to learn all the models that you need to do bilevel planning, this time in BEHAVIOR!
Website:
Code:
Paper:
I'm sure this is not original, but "citation types" would be useful for following related work trails. E.g.,
[1] Directly extending approach
[2] Same problem setting ([2b] used as baseline)
[3] Supports an assertion but otherwise tangential
[4] Unrelated self-citation :)
Looking forward to presenting this work on few-shot imitation learning with programmatic policies at AAAI 2020!
Check out the updated arXiv paper here:
Facebook AI researchers have released PHYRE, a new open benchmark for assessing an
#AI
system’s capacity for reasoning about the physical laws that govern real-world environments.
(2/2) See also (concurrent/independent) work out of Berkeley, Siemens, TUHH:
Very nice results on a real robot.
@jackclarkSF
calls this Franken-RL, which I love.
I think this is some progress, but there is still a lot to figure out, especially:
- How can we learn the symbolic predicates? (c.f. work by George Kondaris; Masataro Asai)
- Can we combine with work on learning behavior priors? (see discussion in NSRT paper)
More to come!
Our “Tools” challenge is finally out! For our
#CogSci2019
paper (with
@realkevinsmith
, arxiv ), we present a fun game to investigate rapid physical trial-and-error learning in humans and machines.
Very excited to share our interactive article:
A visual introduction to Gaussian Belief Propagation!
It's part proposition paper, part tutorial with interactive figures throughout to give intuition.
Article:
Work with:
@talfanevans
,
@AjdDavison
1/n
First we looked at the case where a low-level physics simulator is known and some discrete predicates are defined. We showed how to learn symbolic operators that can be used with a "search-then-sample" TAMP planner.
Paper:
Video:
RL agents explore randomly. Humans explore by trying potential good behaviors, because we have a prior on what might be useful. Can robots get such behavioral priors? That's the idea in Parrot.
arxiv
web
vid
Next we looked at the harder case where we don't have a known simulator. This led us to Neuro-Symbolic Relational Transition (NSRT) models, which can be learned from transition data and used for bilevel TAMP.
Paper:
Video:
@mark_riedl
My understanding is that the subtask graphs are like the recipes, and they learn those from data. Tagging the authors in case they want to clarify :)
@sungryulls
@jaejaywoo
The key idea is that a self-imposed constraint, like "I'm going to stay inside the kitchen", make some parts of a planning problem irrelevant, e.g., the weather outside. So if we can learn a good constraint to impose, we can automatically derive a problem abstraction.