Groups keyboard shortcuts have been updated
Dismiss
See shortcuts
DisneyResearch is at it again (short humanoid)
11 views
Skip to first unread message
Alan Timm
Aug 21, 2024, 10:23:16 AM8/21/24
to RSSC-List
Aw man, you thought BDX was cool, turns out you can use the same motors and build a short king humanoid. low gear reduction bldc gearmotors for everything downstairs, dynamixel type beefy servos for everything else. There's still a lot of promise and potential in reinforcement learning. This policy was trained on a single RTX 4090 consumer GPU over 10 hours.
VMP: Versatile Motion Priors for Robustly Tracking Motion on Physical Characters
"Recent progress in physics-based character control has made it possible to learn policies from unstructured motion data. However, it remains challenging to train a single control policy that works with diverse and unseen motions, and can be deployed to real-world physical robots. In this paper, we propose a two-stage technique that enables the control of a character with a full-body kinematic motion reference, with a focus on imitation accuracy. In a first stage, we extract a latent space encoding by training a variational autoencoder, taking short windows of motion from unstructured data as input. We then use the embedding from the time-varying latent code to train a conditional policy in a second stage, providing a mapping from kinematic input to dynamics-aware output. By keeping the two stages separate, we benefit from self-supervised methods to get better latent codes and explicit imitation rewards to avoid mode collapse. We demonstrate the efficiency and robustness of our method in simulation, with unseen user-specified motions, and on a bipedal robot, where we bring dynamic motions to the real world."
Chris Albertson
Aug 21, 2024, 1:35:16 PM8/21/24
to gestalt73, RSSC-list
I read this a few times. As it turns out every really smart idea is “obvious” AFTER someone else thinks of it. Why is this? Likey fire and the wheel was like that too.
In a nutshell, what I think they are saying is that if “K” is the "kinematic space” and “D” is the "dynamic space”, K+D is very much smaller than K*D. So much smaller that K+D can be trained quickly on consumer-level GPUs.
Typically what people have done is put the full physics based simulation on the computer and then used RL to train it. That takes hundreds of hours but works. During training the robots fall down a lot but after billions of cycles, they get better. What Disney has done here is skipped the physics. Just ignore things like momentum and gravity and train movement. Then turn on the gravity, balance and physics and RL train on “velocity”.
I think what also makes it easier is that the trained motions are imitative, they are created from a human motion. So the simulated robot does not need to spend hours and days to discover that “head-up-feet-down” is the best way to stand.
So they seem to have reduced the cost to train from huge to manageable. Next, let’s do this for the mechanics.
<disneyresearch short king.jpg>--
You received this message because you are subscribed to the Google Groups "RSSC-List" group.
To unsubscribe from this group and stop receiving emails from it, send an email to rssc-list+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/rssc-list/2990cb92-71db-4f92-ae18-c58e72f9b0b8n%40googlegroups.com.
<disneyresearch short king.jpg>
Alan Timm
Sep 17, 2024, 10:16:25 AM9/17/24
to RSSC-List
Robot Motion Diffusion Model: Motion Generation for Robotic Characters
"Recent advancements in generative motion models have achieved remarkable results, enabling the synthesis of lifelike human motions from textual descriptions. These kinematic approaches, while visually appealing, often produce motions that fail to adhere to physical constraints, resulting in artifacts that impede real-world deployment. To address this issue, we introduce a novel method that integrates kinematic generative models with physics based character control. Our approach begins by training a reward surrogate to predict the performance of the downstream non-differentiable control task, offering an efficient and differentiable loss function. This reward model is then employed to fine-tune a baseline generative model, ensuring that the generated motions are not only diverse but also physically plausible for real-world scenarios. The outcome of our processing is the Robot Motion Diffusion Model (RobotMDM), a text-conditioned kinematic diffusion model that interfaces with a reinforcement learning-based tracking controller. We demonstrate the effectiveness of this method on a challenging humanoid robot, confirming its practical utility and robustness in dynamic environments."
Reply all
Reply to author
Forward
0 new messages