Thread by @thomas__lew

📢Excited to share our #ICRA2023 work on robotic table wiping via RL + optimal control!
📖 arxiv.org/abs/2210.10865
🎥
💡RL (for high-level planning) + trajectory optimization (for precise control) can solve complex tasks without on-robot data collection ⬇️

🧽 How can a robot reliably wipe tables to clean spills and crumbs?

The problem is difficult: it requires both high-level planning from image observations and precise low-level control.

We propose an approach in three steps ⬇️

1/3: We "gamify" the high-level wiping problem 🎮: we define an SDE simulator of spills and crumbs dynamics pushed by the robot wiper and an associated reward.

2/3: We train an RL policy in simulation to plan effective high-level wipes from visual observations of spills and crumbs

3/3: We execute the high-level wipes with whole-body trajectory optimization, which allows encoding constraints (avoiding chairs, ...) and accounting for the kinematics of the robot

🤖 We deploy the RL + trajopt approach to hardware and obtain a system that autonomously wipes spills and crumbs. The approach does not require on-robot data collection and transfers zero-shot to hardware

Joint work with an amazing team at @Google Robotics and Everyday Robots! with @Sumeet_Robotics @XiaohanZhang220 @vikassindhwani @Yao__Lu @xf1280 . Special thanks to my internship host Montserrat Gonzalez Arenas for the incredible mentorship last summer!

@Google @Sumeet_Robotics @XiaohanZhang220 @vikassindhwani @Yao__Lu @xf1280 @marioprats @mrcollanderhead @tingnan1986 🦾