←
Tom Silver530 posts
Tom Silver avatar

Tom Silver

@tomssilver

Assistant Professor @Princeton. Developing robots that plan and learn to help people.

◷ Joined October 2011▧ 53 Photos and videos
Tom Silver avatar
Tom Silver@tomssilver·Nov 11

Similar confusion for "agents aren’t closed-loop / reactive." True in (a). But in (b) and (c) the agent outputs a program, and a program can be a closed-loop policy running at control rates.

Tom Silver avatar
Tom Silver@tomssilver·Nov 11

These are very different critiques. (a) is a real latency problem. (b) is a tradeoff that depends on the application. (c) is an amortized cost that competes with engineering or training time.

Tom Silver avatar
Tom Silver@tomssilver·Nov 11

(a) Slow per step: the agent is in the control loop, queried for every action. (b) Slow per task: the agent plans or writes code for each new task, then the robot moves. (c) Slow per domain: the agent writes a policy once, offline, and is not used at deployment.

Tom Silver avatar
Tom Silver@tomssilver·Nov 11

At #CoRL2026 I'm hearing several variations on "agents may be effective at robot decision-making, but they’re too slow" or "not reactive". [E.g., motivating VLAs/WAMs.] I think this means at least 3 different things depending on how the agent is being used. 🧵

Tom Silver avatar
Tom Silver@tomssilver·Nov 1

This week's #PaperILike is "On-line Policy Improvement using Monte-Carlo Search" (Tesauro & Galperin, 1996). A classic example of combined learning and planning, a precursor to AlphaZero (in backgammon), and an easy read. PDF: arxiv.org/pdf/2501.05407

Kartik Garg avatar
⟳ Tom Silver retweeted
Kartik Garg@_pacificax·Oct 27

Sprout Jackson

0:19
Robin Wang avatar
⟳ Tom Silver retweeted
Robin Wang@WangRobin58334·Oct 28

Excited to share UniPred, accepted to the IEEE T-RO 2026(SI on FM for Robotics)! Robots plan long tasks with concepts, but demonstrations only show actions. UniPred learn how to invents those concepts with agent in the loop. 🧵 More in: unipred.github.io

0:45
Abhay Deshpande avatar
⟳ Tom Silver retweeted
Abhay Deshpande@ab_deshpande·Oct 27

DataFarm is cool! Making procedurally generated data look less "robotic" makes a lot of sense. Here's a lightweight JAX/PyRoki port if people are interested, should make it easier to use outside of the TiPToP framework github.com/abhaybd/DataFarm-…

Fine-tuning VLAs on data generated from TAMP seems like an easy win. Turns out it's not that simple! The key issue: distribution mismatch. @samratsahoo_ has a clever workaround: learn to align TAMP to the pretraining distribution. Works really well.

Toru avatar
⟳ Tom Silver retweeted
Toru@ToruO_O·Oct 24

How should we approach research, when the very field we work on is going through rapid and potentially fundamental changes? I recently gave a talk at the Stanford Vision Seminar, reflecting on my PhD journey and how AI is changing the way I think about research. I adapted part of it into a new blog post, organized around three questions: 1. What should we work on? 2. How should we work with AI? 3. What is a PhD for? I hope some of my reflections are helpful. The talk ended with an open discussion, and I’d love to continue it here; curious how others are thinking about these questions. Personally -- as much as there's FOMO, I also think life is very long, and we have time. toruowo.github.io/blog/posts…

Tom Silver avatar
Tom Silver@tomssilver·Oct 20

Adding this to my catalogue of (very) surprising physical strategies discovered by coding agents...

Me: Put the green box on target, clearing obstructions. Astra: You got it, boss 🫡 (Purely python solution generated in a red teamed sandbox using no any other tools, with $20 budget)

0:30