Originally published at: https://developer.nvidia.com/blog/reinforcing-the-value-of-simulation-by-teaching-dexterity-to-a-real-robot-hand/
The human hand is one of the most remarkable outcomes of millions of years of evolution. The ability to pick up all sorts of objects and use them as tools is a crucial differentiator enabling us to shape our world. For robots to work in the everyday human world, the ability to deftly interact with…
What a wonderful work! (“What a time to be alive!” ;-)
Now the proof is made that the training of humanoid robotic hand in simulation (and sim2real) can be carried out on a few GPUs and are no longer reserved for a few happy people who have access to unlimited funds …
The gap between the training time, with and without randomization (for sim2real), seems to be in par (proportionally) with the previous work (cf. OpenAI).
So, if I understood correctly, we should be able to train / try various tasks in fast simulations on a single GPU (as, say: 1-6h / A100 [1], and even 2h / RTX 3090 ! [2]) and then, when we are satisfied, we can train them with randomization with a more powerful workstation.
The article talks about 1.5/2.5 days on AWS g5.48xlarge (8 x A10G GPU) so if things scale linearly it should be possible to train with randomization (sim2real) on a “simple” workstation with 4 x RTX 3080Ti/3090 within 3/5 days… Amazing!
Also, maybe we will see the emergence of a kind of “transfer learning” for humanoid robotic hand…
[1] https://arxiv.org/pdf/2108.10470.pdf , “Isaac Gym: High Performance GPU-Based Physics Simulation For Robot Learning”
[2] ttps://sites.google.com/view/isaacgym
Indeed that’s right, and this is one of the reasons we are so pleased to be able to help democratize this capability.
You can see a demo here of sim2real transfer of a simpler problem (quadruped locomotion) that’s fast enough to be run live on stage in a handful of minutes: Keynote speech Marco Hutter ICRA 2022 - YouTube
Glad you enjoyed our post!
Take care,
-Gav
1 Like
Thanks for the feedback and the interesting link.
Definitely, being able to see the progress of training directly on the robot in “almost” real time is very appreciable.
On this point, I suppose that it is also a question of computing power that you can put. But it should be feasible even for more complex tasks. At least it deserves to be tested.
keep up the good work!-)
I’m looking forward to your code …
Best wishes for the holiday season!