Alumni of Shanghai Jiao Tong University won the best paper, and the awards for CoRL 2022, the top robotics conference, were announced

Since it was first held in 2017, CoRL has become one of the world's top academic conferences at the intersection of robotics and machine learning. CoRL is a single-track conference for robot learning research, covering multiple topics such as robotics, machine learning and control, including theory and applications.

The 2022 CoRL Conference will be held in Auckland, New Zealand from December 14th to 18th.

​

This conference received a total of 504 submissions, and finally accepted 34 Oral papers and 163 Poster papers. The acceptance rate is 39%.

##​

Currently, CoRL 2022 has announced the Best Paper Award, Best System Paper Award, Special Innovation Award, etc. Awards. Kun Huang, a master's degree from the GRASP Laboratory of the University of Pennsylvania and an alumnus of Shanghai Jiao Tong University, won the best paper award at the conference.

Best Paper Award

The winner of the Best Paper Award at this conference is a study from the University of Pennsylvania.

​

• Paper title: Training Robots to Evaluate Robots: Example-Based Interactive Reward Functions for Policy Learning
• Authors: Kun Huang, Edward Hu, Dinesh Jayaraman
##Paper link: https://openreview.net/pdf?id=sK2aWU7X9b8

Abstract: Often, physical interactions help reveal less obvious information, such as we might pull on a table leg to assess Whether it's stable or turning a water bottle upside down to check if it's leaking, the study suggests this interactive behavior could be acquired automatically by training a robot to evaluate the results of the robot's attempts to perform the skill. These evaluations, in turn, serve as IRFs (interactive reward functions) that are used to train reinforcement learning policies to perform target skills, such as tightening table legs. Furthermore, IRF can serve as a verification mechanism to improve online task execution even after full training is completed. For any given task, IRF training is very convenient and requires no further specification.

Evaluation results show that IRF can achieve significant performance improvements and even surpass baselines with access to demos or carefully designed rewards. For example, in the picture below, the robot must first close the door, and then rotate the symmetrical door handle to completely lock the door.

Door locking evaluation example demonstration

The purpose of the following experiment is to combine 3 Visually identical blocks are stacked into a stable tower, with one small block noticeably heavier than the other two, so the best strategy is to place it at the bottom.

Stacked Evaluation Example Demonstration

To check the robustness and generality of the algorithm, this study A D'Claw with 9 joints was used to test it in a real robotic tightening experiment. The purpose of this task is to rotate the 4-prong valve approximately 180° clockwise into the tightened condition (white line on the valve base).

Introduction to the author​

There are three authors who won the CoRL 2022 Best Paper Award this time, namely Kun Huang, Edward Hu, and Dinesh Jayaraman.

Dinesh Jayaraman is an assistant professor at the GRASP Laboratory at the University of Pennsylvania. He leads the Perception, Action, and Learning (PAL) research group, which is dedicated to the intersection of computer vision, machine learning, and robotics. Problem research.

Kun Huang is a master of the GRASP Laboratory at the University of Pennsylvania and studies reinforcement learning under the guidance of Professor Dinesh Jayaraman. He received his BS in Computer Science from the University of Michigan, where he worked on robot perception with Professor Dmitry Berenson. Kun Huang graduated from Shanghai Jiao Tong University with a bachelor's degree. His research interests include robotics and real-world applications. Kun Huang interned at Waymo during his master's degree and will join Cruise as a machine learning engineer after graduation.

LinkedIn homepage: https://www.linkedin.com/in/kun-huang-620034171/

Edward S. Hu I am a doctoral student in the GRASP Laboratory of the University of Pennsylvania, studying under Professor Dinesh Jayaraman. His main research interests include model-based reinforcement learning. Edward received his master's and bachelor's degrees in computer science from the University of Southern California, where he worked on reinforcement and imitation learning in robots with Professor Joseph J. Lim.

Best Paper Shortlist

A total of 3 papers were shortlisted for the Best Paper Award at this conference. In addition to the final winning paper, the other 2 papers are:

##​

• Paper title: Learning Agile Skills via Adversarial Imitation of Rough Partial Demonstrations
• Authors: Chenhao Li, Marin Vlastelica, Sebastian Blaes, Jonas Frey, Felix Grimminger, Georg Martius
• Paper link: https://arxiv.org/pdf/ 2206.11693.pdf

##​

Paper title: Supercharging Imitation with Regularized Optimal Transport
Authors: Siddhant Haldar, Vaibhav Mathur, Denis Yarats, Lerrel Pinto
Paper link: https://arxiv.org/pdf/2206.15469.pdf
Best System Paper Award

The winner of the Best System Paper Award at this conference is a study from CMU and UC Berkeley.

##​

Paper title: Legged Locomotion in Challenging Terrains using Egocentric Vision

• Authors: Ananye Agarwal, Ashish Kumar, Jitendra Malik, Deepak Pathak
• Paper link: https://arxiv.org/pdf/2211.07638.pdf

Abstract:

Animals are able to use vision to make precise and agile movements, and replicating this ability has been a long-standing goal of robotics. The traditional approach is to break the problem down into an elevation mapping and foothold planning phase. However, elevation mapping is susceptible to glitches and large-area noise, requires specialized hardware and is biologically infeasible. In this paper, researchers propose the first end-to-end locomotion system capable of traversing stairs, curbs, stepping stones, and voids in a medium-sized building using a single facade. The results are demonstrated on a quadruped robot with a depth camera. Due to the small size of the robot, there is a need to discover specialized gait patterns not found elsewhere. The camera needs to master the strategy of remembering past information to estimate the terrain behind and below it.

The researchers trained the robot’s strategy in a simulated environment. Training is divided into two stages: first using reinforcement learning to train a policy on deep image variants with low computational cost, and then refining it to a final policy on depth using supervised learning.

​

The resulting strategy is transferable to the real world and can be run in real time on the limited computing power of the robot. It can traverse a wide range of terrains while being robust to disturbances such as slippery surfaces and rocky terrain.

Stepping Stones and Gaps

The robot is able to step over bar stools in various configurations and adjust the step length to Excessive gap. Since there are no cameras near the back feet, the robot has to remember the position of the bar stool and place its back feet accordingly.

Stairs and curbs

The robot is able to climb up to 24 cm High, 30 cm wide stairs. Strategies apply to different stairways and curbs under various lighting conditions. On unevenly spaced stairs, the robot will initially get stuck, but will eventually be able to use climbing behavior to cross these obstacles.

Unstructured terrain

Robots can traverse areas that are not part of their training One of the categories, unstructured terrain, shows the generalization ability of the system.

Movement in the Dark

The depth camera uses infrared light to project patterns , accurately estimating depth even in almost no ambient light.

Robustness

Strategy to high forces (throwing 5 kg weight from height) and slipperiness The surface (water poured on plastic sheet) is robust.

Author introduction​

This item The study has four authors.

Jitendra Malik is currently the Arthur J. Chick Professor in the Department of Electrical Engineering and Computer Science at UC Berkeley. His research areas include computer vision, computational modeling of human vision, computer graphics and biology. Image analysis, etc.

Ashish Kumar, one of the authors of this award-winning study, is his doctoral student.

Deepak Pathak is currently an assistant professor at Carnegie Mellon University. He received his PhD from the University of California, Berkeley, and his research topics include machine learning, robotics and computer vision.

Ananye Agarwal, one of the authors of this award-winning study, is his doctoral student.

In addition, Deepak Pathak has another study on the shortlist for the Best System Paper Award at this conference.

• Paper title: Deep Whole-Body Control: Learning a Unified Policy for Manipulation and Locomotion
• ##Authors: Zipeng Fu, Xuxin Cheng, Deepak Pathak
• Paper link: https://arxiv.org/abs/2210.10044

Special Innovation Award

This time The conference also selected a special innovation award. This research was jointly completed by many researchers from Google.

##Paper title: Do As I Can, Not As I Say: Grounding Language in Robotic Affordances
Authors: Brian Ichter, Anthony Brohan, Michael Ahn, etc.
Paper link: https://arxiv.org/pdf/2204.01691.pdf

Paper abstract: Large language models can encode a large amount of semantic knowledge about the world, and such knowledge is very useful for robots. However, language models have the disadvantage of lacking experience with the real world, which makes it difficult to leverage semantics to make decisions on a given task.

Researchers from Google propose to provide a real-world foundation for large language models through pre-training skills that are used to constrain the model to come up with natural language that is both feasible and contextually appropriate. operate. Robots can serve as the "hands and eyes" of language models, which provide high-level semantic knowledge about the task. This study shows how low-level skills can be combined with large language models so that the language model provides high-level knowledge about the processes that execute complex and time-extended instructions, while the value functions associated with these skills provide the means to connect this knowledge to specific physical environments. required basis.

The researchers used this principle when combining a large language model (LLM) with the physical tasks of the robot: in addition to letting the LLM simply interpret an instruction, you can also use It evaluates the probability that a single action will help complete the entire high-level instruction. Simply put, each action can have a language description, and we can use the prompt language model to let it score these actions. Furthermore, if each action has a corresponding affordance function, it is possible to quantify its likelihood of success from the current state (e.g., a learned value function). The product of two probability values ​​is the probability that the robot can successfully complete an action that is helpful to the instruction. Sort a series of actions according to this probability and select the one with the highest probability.

The example below shows a robot helping to get an apple:

The above is the detailed content of Alumni of Shanghai Jiao Tong University won the best paper, and the awards for CoRL 2022, the top robotics conference, were announced. For more information, please follow other related articles on the PHP Chinese website!