Assistant professor, @Cornell_CS. Learning, perception, and control for building more dynamic robots.
Cambridge, MA
Joined August 2019
- Tweets37
- Following139
- Followers393
- Likes66
Pinned Tweet
I'm so excited to announce I will be joining @Cornell_CS as an assistant professor next fall!
I can't wait to launch my lab and start building robots with human-level dexterity and adaptability. 🤖🏃🤹
ALT Preston sitting at a table with two robots, signing their offer letter from @Cornell_CS.
We got the humanoid recipe (human demo → retarget → RL) to work for dexterous hands. Key idea: use interaction-preserving retargeting, not just IK.
@yunhai got tool use (including scissors!) on real hands from one human demo. Good retargeting → surprisingly natural motion.
Can robots learn contact-rich tool use, like operating a pair of scissors✂️ or turning a screwdriver🪛, from human data?
Introducing REGRIND: a minimalist retargeting-guided RL recipe for dexterous manipulation.
🌐 yunhaifeng.com/REGRIND/
📄 Paper + code on the site!
🧵👇
So cool to see the loco-manipulation stack we developed at RAI out in the wild! Congrats to @simonlc_ and the team for the awesome demo.
This ability to adapt RL policies using test-time reward spec + MPC/search is really cool / basically why we built judo.
See Spot perform dynamic whole-body manipulation. Using a combination of reinforcement learning (RL) and sampling-based control, the robot is able to autonomously drag, roll, and stack tires weighing 15 kg (33 lb), well above its maximum arm lift capacity.
Learn more about coordinating locomotion and manipulation processes: rai-inst.com/resources/blog/…
If you haven't already, you should check out judo, the MPC package underlying a lot of the skills here (including the tire flip-up!!) - we're excited to see what other cool ideas people can cook up
github.com/bdaiinstitute/jud…
Preston Culbertson retweeted
Just over one week until registration closes for NERC! Register by next Friday, Oct 3, to join us in Ithaca next month.
🎟️events.ces.scl.cornell.edu/e…
Preston Culbertson retweeted
Just one month until NERC 2025!
Come to Cornell on Oct 11 for a full day of cutting-edge robotics research and community.
📍 Cornell University, Ithaca, NY
🗓️ Oct 11, 2025
🎟️ Register by Oct 11: events.ces.scl.cornell.edu/e…
Join us in Ithaca this October for NERC!
Submissions are due this Tuesday, Sept 2 - excited to see everyone at Cornell in a few weeks.
Reminder: Poster and Rising Star submissions for NERC 2025 are due next Tuesday, Sep 2!
Submit now for a great opportunity to share your research with robotics faculty and peers.
🔗 openreview.net/group?id=NERC…
📅 Deadline: Sep 2
🥋 We're excited to share judo: a hackable toolbox for sampling-based MPC (SMPC), data collection, and more, designed to make it easier to experiment with high-performance control.
Try it: pip install judo-rai
There’s more to come:
⚡ GPU-accelerated sim (via mujoco_warp)
🦾 New tasks: locomotion, dexterous manipulation
🧠 Hooks for adding learning-based components
judo is still early days, but we hope it lowers the barrier to SMPC and sparks new directions in control and robot learning.
Try it here 👉 github.com/bdaiinstitute/jud…
or: pip install judo-rai
Congrats to @albert_h_li for winning Outstanding Paper today at the CoRL workshop for learning fine + dexterous manipulation!
ICYMI: For #CoRL2024 we released a dataset of 3.5M (!) dexterous grasps, with multi-trial labels and perceptual data for 4.3k objects.
Our takeaways: scale matters, and refining grasps > better sampling. Hoping our data can enable more vision-based grasps in hardware!
There have been many recent big grasping datasets, but few demos of real-world grasping using generative models. How do we achieve this?
Introducing: Get a Grip (#corl2024)!
We show that instead of generative models, discriminative models can attain sim2real transfer!
👀🧵👇