Interested in making your bipedal robots to be athletes? We summarized our RL work to create robust & adaptive controllers for general bipedal skills. 400m-dash, running over terrains/against perturbations, targeted jumping, compliant walking, not a problem for bipeds now.🧵👇
Our goal is to push the limits of the agility of legged robots. Using a single RL-based multi-task policy, a biped Cassie learns to perform various challenging jumps, including standing long jump, high jump, and/or multi-axes jumps, with significant robustness, in the real world.
Can a quadruped perform agile and dynamic locomotion with fast and precise manipulation (like ball interception) skills? We create a robotic goalkeeper using Mini Cheetah with deep RL and it shows an 87.5% goal save ratio! More details can be found in our recent paper and video.
I think the coolest thing of GenLoco is that, if there is a new quadrupedal robot developed, we can just download the pre-trained model and use it as a baseline controller for the new robot without worrying about developing a new controller from scratch! Welcome to try it out!
Can we train a *single* policy that can control many different robots to walk? The idea behind GenLoco is to learn to control many different quadrupeds, including new ones not seen in training.
Code
Video
I'm very excited to share our recent work of using RL for a versatile and robust walking policy on a bipedal robot Cassie!
The video is here: . The paper is at .
Huge thanks to
@whshkan
@xbpeng4
@GlenBerseth
@pabbeel
@svlevine
!
Super exciting to see more people using animation for legged robots! Back in 2020, my first paper was to use animation to draw expressive motions for Cassie. At that time, RL on real bipeds was still a question (we used traj opti)! Now, we also used anm for RL jumping on Cassie!
This new robotic character from
@Disney
Research is ADORABLE, and also demonstrates a new pipeline for deploying expressive behaviors that are robust enough to work in the real world. We talk to the Disney team about how it all works:
#IROS2023
Quadruped robots already have four manipulators (aka legs), do we really need extra arms for them? 🤔
Dive into our latest work: quadruped robots handle complex loco-manipulation tasks autonomously with just legs! 🐾
The takeaway is: the multi-task policy improves the robustness. By learning from various jumping tasks, the robot can generalize the learned tasks, and pick the best to recover from poor leaning pose or perturbation.
Paper:
Video:
We found that RL can bring robustness and adaptivity for dynamic & real-time bipedal control, but only with a proper formulation. Using a long history of robot I/O, complementing it with a short history, training it end-end, and adding task randomization, are the recipes.
We conducted extensive benchmark and ablation studies in both sim and real to support this. Details are in the paper:
Summary video:
This work cannot be done without
@xbpeng4
@GlenBerseth
@svlevine
@pabbeel
& Koushil
We decompose the goalkeeping task into 2 subtasks using RL: manipulation planning and locomotion control for multiple dynamic skills. This work is with X. Huang and a team of undergrads, and
@xbpeng4
and K. Sreenath.
Paper:
Video:
btw, a more general takeaway from GenLoco is, even when we are focusing on developing robot-specific controller using RL, randomizing robot morphology/kinematics can help to improve the robustness of the policy in sim/sim2real.
Not the first time using RL for locomotion control on bipeds but one of the first time using RL to directly output desired joint positions to control while showing great robustness! The cover figure is from our previous animation work though--this is another interesting paper😉
Introducing 𝐀𝐋𝐎𝐇𝐀 𝐔𝐧𝐥𝐞𝐚𝐬𝐡𝐞𝐝 🌋 - Pushing the boundaries of dexterity with low-cost robots and AI.
@GoogleDeepMind
Finally got to share some videos after a few months. Robots are fully autonomous filmed in one continuous shot. Enjoy!
Happy to share a recent RAL-IROS paper on using multiple quadrupeds to collaborate and tow a load using cables, which can be slack/taut, while navigating in cluttered spaces, fully autonomously and online. A cool demo video is here:
Glad to share a recent paper that uses RL to enable a quadruped to precisely shoot a soccer ball to random targets. We first train a robust shooting controller then train and finetune a motion planner. This paper is in the IROS Best RoboCup Paper finalist this year!
@DrJimFan
Thanks! The paper is still in the submission so it is not opensourced. But we do have the plan to opensource the code after everything is wrapped up.
New experiment paper using quadrepeds to autonomously explore unknown environments with a skill of jumping through the obstacles that is not traversable by walking only.
Glad to share our recent work including collaborators
@ZhongyuLi4
and
@ayush1292
about quadrupedal robots. The idea of this paper allows the robots to switch behaviors between walking around obstacles or jumping through them in order to reach its goal for a navigation problem.
I want to share our recent work accepted to ACC about control barrier functions, where the hyperparameter is optimized to ensure point-wise feasibility.
paper link:
github repo:
Hope this method saves people from excessive tuning...
@JameScottX
Thanks for your interest! The freq is 33Hz, and yes, Cassie is a torque-controlled robot. Happy to chat more and you can find more details in our paper:
Glad to share our recently submitted work on a bipedal robot Cassie autonomously navigating unknown environments with height constraints. The entire autonomy runs in real-time and real-world!
Paper:
Video: