Learning for Agile Robotics

Learning for Agile Robotics Workshop at CoRL 2022

Awards

Best Paper:

Learning to Navigate Over Clutter in Indoor Environments using Vision (poster)
Simar Kareer, Naoki Harrison Yokoyama, Dhruv Batra, Sehoon Ha, Joanne Truong

Honorable Mention:

Agile Catching with Whole-Body MPC and Blackbox Policy Learning
Anish Shankar, Stephen Tu, Deepali Jain, Sumeet Singh, Krzysztof Marcin Choromanski, Saminda Wishwajith Abeyruwan, Alex Bewley, David B D'Ambrosio, Jean-Jacques Slotine, Pannag R Sanketi, Vikas Sindhwani

We have seen tremendous progress and successes in applying learning to robotics over the last decade. Recently, learning based approaches have emerged for developing dynamic robots such as quadrupeds, ping-pong and drones. However, learning based robots are yet to demonstrate the capability to be as agile as humans or animals, like the traditionally nonlinear optimization based controls based agents have. Learning agile skills needs to overcome challenges such as modeling and adaptation in fast changing environments, low latency and high frequency perception and control, larger sim-to-real gap, need for wider safety margins, operating at hardware limitations and many more.

In this workshop, we plan to invite researchers working on making robots drive fast, run, fly, play sports, catch, juggle, etc. using machine learning to share their experience. The goals of the workshop include:

  • Fostering collaboration between a diverse group of researchers and practitioners working on a wide range of agile robotic domains (e.g. agile locomotion, ground vehicles, drones, ping-pong, catching, juggling, etc.).

  • Understanding the current limitations and inefficiencies in our methods and how modern ML architectures (Transformers, graph NNs and recurrent neural networks for example) and approaches such as generative modeling (e.g. GANs, VAEs, AR models, diffusion models, etc.), unsupervised learning, meta-learning, representation learning, domain adaptation, offline RL, etc. can be used to overcome the challenges in agility.

  • Learning common (or contrasting!) ML ideas across traditional robotics approaches and modern ML based approaches and as well as across diverse application domains; spanning various technical topics such as sensor fusion, algorithms (perception, learning, planning, state estimation, control) and systems and how it affects the agent learning.

  • Raising awareness amongst the robot learning community about the rich set of problems in learning for agile robotics.

Accepted Papers

Click on the links below to access the PDFs via OpenReview.

Legged Locomotion in Challenging Terrains using Egocentric Vision
Ananye Agarwal, Ashish Kumar, Jitendra Malik, Deepak Pathak

Utilizing Human Feedback for Primitive Optimization in Wheelchair Tennis
Arjun Krishna, Zulfiqar Haider Zaidi, Letian Chen, Rohan R Paleja, Esmaeil Seraj, Matthew Gombolay

Learning Agile Paths from Optimal Control
Alex Beaudin, Hsiu-Chin Lin

Fearful Goal Generation for Reliable Policy Learning (direct PDF link)
Charlie Gauthier, Florian Golemo, Glen Berseth, Liam Paull

Safe Inverse Reinforcement Learning via Control Barrier Function
Yue Yang, Letian Chen, Matthew Gombolay

A Hierarchical Reinforcement Learning Approach to Control Legged Mobile Manipulators
Florian Golemo, Simon Chamorro, Martin Weiss, Liam Paull, Christopher Pal

Perception-Based Rewards for Robotic Shepherd Teams Maneuvering Split Flocks
John Baxter, Yazied Abdulhamid Hasan, César Alejandro Salcedo, Lydia Tapia

Agile Catching with Whole-Body MPC and Blackbox Policy Learning
Anish Shankar, Stephen Tu, Deepali Jain, Sumeet Singh, Krzysztof Marcin Choromanski, Saminda Wishwajith Abeyruwan, Alex Bewley, David B D'Ambrosio, Jean-Jacques Slotine, Pannag R Sanketi, Vikas Sindhwani

FlowDrone: Wind Estimation and Gust Rejection on UAVs Using Fast-Response Hot-Wire Flow Sensors
Nathaniel Simon, Allen Z. Ren, Alexander Piqué, David Snyder, Daphne Barretto, Marcus Hultmark, Anirudha Majumdar

Learning to Navigate Over Clutter in Indoor Environments using Vision (poster)
Simar Kareer, Naoki Harrison Yokoyama, Dhruv Batra, Sehoon Ha, Joanne Truong

Learning Agile Skills via Adversarial Imitation of Rough Partial Demonstrations (poster)
Chenhao Li, Marin Vlastelica, Sebastian Blaes, Jonas Frey, Felix Grimminger, Georg Martius

Invited Speakers


(alphabetical order)

David B. D’Ambrosio - Google

Scott Kuindersma - Boston Dynamics

Laura Graesser - Google

Guanya Shi - Caltech/CMU

Jitendra Malik - UC Berkeley

Hae-Won Park - KAIST

Jan Peters - TU Darmstadt

Amirreza Shaban - U Washington

Anima Anandkumar - Caltech

Schedule

Morning

8:00-8:25: Breakfast

8:25-8:30: Introduction

8:30-9:00: Keynote #1: Scott Kuindersma (Boston Dynamics)

9:00-10:30: Short talks session #1

9:00 - 9: 15: Matthew Gomblay (Gatech): Towards Robo Federer: Agile, Safe, and Cognitive Robot Systems

9:15 - 9:30: Guanya Shi (incoming CMU): Neural-Control Family: Safe Agile Deep-learning-based Control in Dynamic Environments

9:30 - 9:45: Giuseppe Loianno (NYU): Learning Safe, Adaptive, and Agile Flight Control

9:45: 10:00: Johannes Betz (UPenn): Autonomous Vehicles on the Edge: Learning to drive fast at the vehicle dynamics limits

10:00 - 10:15: Xuesu Xiao (GMU): Learning Agile Ground Maneuvers in Highly Constrained and Off-Road Conditions

10:15 - 10:30: Laura Graesser & David B. D’Ambrosio (Google): When you’re forced to move fast: lessons learned from robotic table tennis

10:30-11:00: Coffee break and posters

11:00-11:30: Keynote #2: Jan Peters (TU Darmstadt)

11:30-12:15: Discussion #1: "Main obstacles and challenges for agile robotics, barriers in applying robot learning to agile robotics".

12:15-13:30: Lunch break


Afternoon

13:00-14:00: Posters and demos

14:00-14:30: Main CoRL Opening session

14:30-15:00: Keynote #3: Anima Anandkumar (Caltech)

15:00-15:30: Coffee break, posters and demos

15:30-17:00: Short talks session #2

15:30 - 15:45: Pulkit Agrawal (MIT): Rapid Contact-Rich Control via Reinforcement Learning

15:45 - 16:00: Heni Ben Amor (ASU): Phase Estimation for Agile and Interactive Robotic Skills

16:00 - 16:15: Hae-Won Park (KAIST): Agile Legged Robot Locomotion with Model Predictive Control and Reinforcement Learning

16:15 - 16:30: Jitendra Malik (UC Berkeley): Vision and Agile Walking

16:30 - 16:45: Amirreza Shaban & Xiangyun Meng (U Washington): Perception for Off-Road Autonomous Driving

16:45 - 17:00: Andrew Razjigaev (QUT): End-to-End Design of Bespoke, Dexterous Snake-Like Surgical Robots

17:00-17:05: Closing remarks and announcements. Overflow poster session time.


Evening

Closing social event

Location

The Owen G. Glen Building / Business School of the University of Auckland
Building 260 of the University of Auckland, 12 Grafton Rd, Auckland CBD – main venue
abbreviated as OGG

Building 260
Room 040 (260.040)
Team Based Learning Lab
Capacity: 120

Technical Support: Joao Buzzatto

See https://corl2022.org/workshops/ for additional details.

Remote Attendance

For Zoom info, please refer to the conference registration or contact the workshop organizers.

Organizers


Alphabetical order, please contact the organizers at organizers@agilerobotscorl2022.com.

Anima Anandkumar - Caltech

Ken Caluwaerts - Google

Alejandro Escontrela - UC Berkeley / Google

Chuchu Fan - MIT

Ken Goldberg - UC Berkeley

Laura Graesser - Google

Atil Iscen - Google

Chase Kew - Google

Pannag Sanketi - Google

Davide Scaramuzza - University of Zurich

Jie Tan - Google

Sarah Tang - Waymo

Xuesu Xiao - George Mason University / Everyday Robots

Yuxiang Yang - University of Washington

Wenhao Yu - Google

Tingnan Zhang - Google

Call for Papers

Learning based robots have yet to demonstrate the capability to be as agile as humans or animals. Learning agile skills necessitates overcoming challenges such as modeling and adaptation in fast changing environments, low latency and high frequency perception and control, large sim-to-real gaps, a need for wider safety margins, operating at hardware limitations and more.

In the CoRL 2022 Workshop on Learning for Agile Robotics, we invite researchers using machine learning to make robots move fast to share their research.

Topics of interest include but are not limited to:

  • Robots that run, jump, drive, fly, play sports, catch, juggle, dance, ....

  • Sensor fusion for fast movements

  • System design for agile robots

  • Algorithms (e.g. perception, learning, planning, state estimation, control)

  • Work that sheds light on current limitations and inefficiencies in learning for agile robotics and how modern ML architectures and approaches can be used to overcome these challenges.


Submission process

Submission website: OpenReview: CoRL Agility Workshop


Submissions should use the CoRL 2022 template and be 4 pages (plus as many pages as needed for references and appendices, reviews will be based on the main paper).
Following the main CoRL conference we use OpenReview and the review process will be
double blind.


Accepted papers and eventual supplementary material will be made available on the workshop website. However, this does not constitute an archival publication and no formal workshop proceedings will be made available, meaning contributors are free to publish their work in archival journals or conferences.


Note that we will accept submissions accepted to other conference or journal proceedings at the time of submission.


A note on demos: We encourage participants to also bring their robots for demos. Demo sessions are organized by the main CoRL committee. See https://corl2022.org/call-for-demos/ for additional information.


Review criteria

We will reject submissions that are not focused on agile robot learning.

Important Dates

Submissions opening: September 7, 2022
Demo application deadline (via main conference): September 20th, 2022

Submission deadline: October 16, 2022 Extended: October 28, 2022 (midnight AoE)
Notification of acceptance: October 28, 2022 November 17, 2022
Date of Workshop: December 15, 2022 (full day)

FAQ

  • Which poster session am I presenting at?

There is no difference between poster sessions as we will have enough poster boards. Authors are asked to have their poster ready by the first coffee break.

  • What is the poster size?

The recommend poster size is A1. We will not have TV screens available.

  • Can posters being setup by other people?

Yes, posters can be setup by other people. However, it is recommended that one of the authors be physically at the conference, since this year's CoRL conference has been optimized for an in-person experience.

  • How can I attend remotely?

We will post information closer to the workshop (Zoom link/PheedLoop).

  • Where can I print a poster locally?

One option is James Ashford in Auckland.

  • Will monitors be available for the poster sessions?

Yes. We will have monitors available if you wish to use them. This is optional and can be used to display videos of your robots in action in addition to your poster. Note that for workshop poster/papers we only support on site presentations, so you will not have Zoom links for the posters.

Sponsors