Thread by @drfeifei

V. proud that our @StanfordSVL has 9 papers accepted by #CoRL2021! “Seeing is for doing.” Computer vision & robotic learning are deeply intertwined areas of research as we work to understand and build active & learning AI agents to interact in &w/ the real world. Enjoy below:) 1/

iGibson2.0 is a robotic learning environment that enables next generation simulation and benchmarking of robotic learning, and human-robot interaction tasks. 2/

https://twitter.com/drfeifei/status/1425204811771744263?s=20

iGibson2.0 enabled the development of BEHAVIOR, a benchmark dataset & environment for complex robotic learning tasks. Today's algorithms performs near ZERO when doing these 100 real-world inspired tasks (Tab2). Very excited to see more progress to come! 3/

https://twitter.com/drfeifei/status/1431280905436368896?s=20

A critical element of real-world interaction is 3D scene reconstruction. This paper shows how to infer a fully editable 3D scene from a single image! 4/

https://twitter.com/ZakharovSergeyN/status/1438592544644861959

Vision is not the only sensory modality for #AI agents to understand the world. This paper uses sound to infer object properties! 5/

https://twitter.com/SamuelPClarke/status/1439865499689439236

ObjectFolder: Combining vision, sound & touch, we present a dataset of 100 objects w/ their implicit neural representations that encode these multi-modal dimensions for downstream tasks,e.g. object recognition, robotic grasping, cross-sensory retrieval. 6/

https://twitter.com/RuohanGao1/status/1438713307959943171

Learning from humans to enable robotic tasks such as mobile manipulation is very useful. Here we present a novel error-aware imitation learning method via a more intuitive teleoperation system called MoMaRT. 7/

https://twitter.com/josiah_is_wong/status/1440163888914333706

More learning from humans by robots! This time is to learn human-human collaboration to enable human-robot collaboration behaviors. Our novel algorithm called Co-GAIL combines imitation learning with a co-policy optimization process to achieve that. 8/

https://twitter.com/chenwang_j/status/1438687048504786945?s=20

The Robomimic framework offers a suit of tasks, large-scale human demo datasets collected on robot manipulation domains, & policy learning algorithms. This work is selected as an Oral Presentation #CoRL2021 9/

https://twitter.com/AjayMandlekar/status/1425140033804070916

EMBR-Example-driven model-based RL-takes a step forward towards learning long-horizon manipulation tasks via visual & motor learning! Check out our results of a real robot arm organizing (ikea!) desk & drawer , or put away cluttered markers and more! 10/

https://twitter.com/bohanwu_1/status/1440707067195494415?s=20

Too many awesome students, faculty and collaborators to acknowledge in one tweet. Very grateful to be part of such an energetic, creative, supportive and fun research community @StanfordSVL END/ svl.stanford.edu