Unlike humans who favor a dominant hand for fine dexterous skills, robots should execute ambidextrous manipulation with equal proficiency. Introducing SYMDEX, our algorithm for Ambidextrous Bimanual Manipulation leveraging robot’s inherent bilateral symmetry as an inductive bias.

🧩Web: https://t.co/kUplDJIjOg, 📷Paper: https://t.co/ypyoHzmRWs. It is a pleasure to work with my amazing collaborators Yufeng Jin, @OrdonezApraez and supervisors @GeorgiaChal, @liu_puze, and Claudio Semini.

Notably, our algorithm naturally scales to multi-arm setup with more complicated symmetry groups!

We further distill them into a global policy that is independent of the hand-task assignment and allows smooth sim-to-real transfer.

SYMDEX frames complex bimanual manipulation as a multi-task multi-agent problem and learns dedicated equivariant policies for each subtask.

Our method significantly outperforms other diffusion-based methods on high-dimensional control benchmarks, and is also competitive with state-of-the-art non-diffusion based RL methods while requiring fewer algorithmic design choices and smaller update-to-data ratios!

Meet our first general-purpose robot at @DexmateAI https://t.co/jFabzsDSHJ Adjustable height from 0.66m to 2.2m: compact enough for an SUV, tall enough to reach those impossible high shelves. Powerful dual arms (15lbs payload each) and omni-directional mobility for ultimate

Grateful to have worked with my awesome collaborators Rickmer Krohn, @taochenshh, @aajay3110, and supervisors @GeorgiaChal, @pulkitology!

We design a series of new challenging robot navigation and manipulation tasks with a high degree of multimodality, serving as a testbed for multimodal policy learning. (5/5) 🧩Web: https://t.co/akgO1y7Egr. 📔Paper: https://t.co/YxchrzotWy.

Moreover, we achieve explicit mode control by policy conditioning on a mode-specific latent embedding during training, shown to be beneficial for online replanning. (4/5)

Second, the greediness of RL objective will collapse the policy into a single mode. We address this issue by using mode-specific Q-functions and constructing multimodal batch for policy learning. (3/5)

First, we explicitly discover behavior modes via an off-the-shelf unsupervised hierarchical clustering approach. Different modes should act differently at certain states and then diverge, therefore being distinguishable from trajectories/sequences of states. (2/5)

How can an RL agent successfully solve a task while showcasing versatility of behaviors—a property intuitive to intelligent systems like humans? Introducing Deep Diffusion Policy Gradient (DDiffPG) - our new algorithm for learning multimodal behaviors from scratch! (1/5)

Robot learning in the real world can be expensive and unsafe in human-centric environments. Solution: Construct simulation on the fly and train in it! Excited to share RialTo, led by @marceltornev on learning resilient policies via real-to-sim-to-real policy learning! A 🧵 (1/12)

Working with massive parallel simulation like Isaac Gym? Wonder how to learn policies faster and better than PPO? Check out our poster #102 (1:30pm on Thur) #ICML2023 . We have also open sourced the code. 📔Paper: https://t.co/a5iNnjCKBl 🧑‍💻Code:

Introducing Decision Diffuser, a conditional diffusion model that outperforms offline RL across standard benchmarks – using only generative modeling training! Decision Diffusers can also combine multiple constraints and skills at test-time. Website: https://t.co/bQErTTKHEc 1/5

“ElegantRL: Much Much More Stable than Stable-Baseline3” by Xiao-Yang Liu https://t.co/CZMMEwO1Hx

Check out ElegantRL, a lightweight, stable, and efficient deep reinforcement learning library with < 1000 lines of code designed to make our life easier! It also powers FinRL ( https://t.co/kshYms4sVR). https://t.co/ijpwQijcbH