📝 The graduate admission application is open! 📆 Deadlines: ✔️ PhD: 12/1/21 ✔️ MIMS: 1/6/22 ✔️ 5th Year MIDS: Early (app fee waived!) - 11/4/21; Final - 3/2/22 ✔️ MIDS & MICS: rolling throughout the year! Grad app: bit.ly/3zsoOVZ

“…What started in 2017 as an email discussion and later a Facebook Group has grown into a global movement of 3,800 members in more than 50 countries. Black in AI works in academics, advocacy, entrepreneurship, financial support, and summer research programs.”

'basketball_in_hoop'; one of many new tasks joining the #RLBench family of 100+ tasks in V1.2. Coming early November! 🤖 RLBench is still the hardest manipulation sim-benchmark to date due to its large-scale focus on vision, sparse rewards, and multi-stage tasks.

Great paper from our lab on grasping transparent objects will be presented at Conference on Robot Learning in London next month

It’s notoriously difficult to model the mechanics of compliant robot jaw tips during grasping! We found that a new tool from computer graphics can help. IPC-GraspSim, from AUTOLab UC Berkeley. Paper, data, video: sites.google.com/berkeley.edu/i… (1/9)

An inspiring talk by Ken Goldberg from Berkeley Engineering on robot assisted surgery. Imperial College London

Why is generalization hard in RL? Can "just adding more data" fix it? Turns out that in general, the answer is no. In a new blog post, Dibya Ghosh discusses this question: bair.berkeley.edu/blog/2021/11/0… Trying to generalize induces a POMDP, even if the problem is an MDP!

Fun to be in London with Jeff Ichnowski Ashwin Balakrishna Brijen Thananjeyan Ryan Hoque Yahav Avigal @AleEscontrela & Justin Kerr from AUTOLab checking out Anička Li’s elegant floating robots Tate, eating Yorkshire pudding and looking fwd to robotics rendezvous this week:

Come check out our poster (3-15) at #CoRL2021 tomorrow evening!

Planar Robot Casting for deformable materials aims to achieve a desired final state from one dynamic launching action. Our work from AUTOLab @Berkeley_AI learn it using a self-supervised “Real2Sim2Real” framework. Data, paper, and presentation: tinyurl.com/robotcast (1/8)

Unlike fly-fishing, shuffleboard, and bowling, Planar Robot Casting allows self-supervised learning. This real-world robot control problem includes nondeterminism, surface friction, and deformable materials, and it's relatively easy to set up; hoping others study it also:

Robots can be designed to be symbiotic with workers, freeing us to focus on the myriad of tasks AI can’t do well. Proud that this is an axiom for AmbiRobotics.

Thank you ini_fadelli for a great summary of our work on this new robotics problem with Toyota Research Institute (TRI) AUTOLab techxplore.com/news/2021-11-r… TechXplore

Can a robot teach itself to grasp complex objects? Learned Efficient Grasp Sets (LEGS) can help robots efficiently learn to grasp novel, out-of-distribution objects. Research from AUTOLab UC Berkeley. Paper, website: sites.google.com/view/legs-exp-… (1/8)

Interested in safe and robust learning for control? Come check out our NeurIPS 2021 Workshop on Safe and Robust Control of Uncertain Systems (Website: sites.google.com/view/safe-robu……) on 12/13 from 8 AM - 4 PM PST! You can register here: neurips.cc.

Mark your calendars for this NeurIPS workshop on 12/13

Kudos to Obvious Ventures for demonstrating something far from obvious: progressive values and profitable ventures can be highly compatible.

JAXopt v0.2 is out! github.com/google/jaxopt/… The main highlight of this release is an implementation of OSQP, a GPU-friendly quadratic programming solver. Our implementation of course supports implicit differentiation ;) Thanks to our intern Louis Béthune for the hard work.

Very cool work on leveraging German Shepherd mocap data to learn style rewards!