Ville🤖 (@VilleKuosmanen) / X

Ville🤖

2,021 posts

Ville🤖

@VilleKuosmanen

gentleman scientist 🤖 @voyagerobotics

London, UK

Joined May 2020

Pinned
Ville🤖
@VilleKuosmanen
Apr 23
A month ago @pravsels and I set out to reproduce @physical_int’s RL Token paper. Today, I am sharing our research notes from the journey we’ve been on. x.com/VilleKuosmanen…
Ville🤖
@VilleKuosmanen
Mar 24
"RL Token" looks like a great and surprisingly simple post-training methodology for optimising robot models for dexterous tasks in the real world! Over the next few weeks, me and @pravsels will be attempting to reproduce the results (& open source the code) Stay tuned 👀
17K
Ville🤖
@VilleKuosmanen
Feb 23, 2025
After fixing a few bugs with inference code, I finally have a working pi0 set up! Fine-tuned overnight on single task data, it learns rough controls and no longer outputs unsafe actions. But mastering the task and form factor will require much more compute. Stay tuned!
00:00
34K
Ville🤖
@VilleKuosmanen
Jun 2, 2025
I wrote a Manifesto for open-source robotics! If you are passionate about the topic or want to learn about the state of open-source in today's robotics and physical AI, give it a read 🧵
20K
Ville🤖
@VilleKuosmanen
Feb 10, 2025
A few of you have asked me about my experience with pi0… I fine-tuned the model overnight on a 4090. It failed to learn anything useful and was just outputting non-compliant actions my safety system stopped from breaking the robot (pic related)
56K
Ville🤖
@VilleKuosmanen
Mar 10, 2025
I rented an H100 and fine-tuned @physical_int ‘s pi0 overnight. We’re still far from fully utilising the potential of this model, but it produced my first successful rollout!
00:00
66K
Ville🤖
@VilleKuosmanen
Apr 3, 2025
Pi0 with Gemma 3 chains of thought now working pretty reliably!
00:00
23K
Ville🤖
@VilleKuosmanen
May 21, 2025
Do AI robots see the world like we do? I dove head first into latent space to uncover the attention maps that show how my robot sees and understands the world.
00:00
44K
Ville🤖
@VilleKuosmanen
Aug 15, 2025
Quick proof of concept for teleop with offsets between leader and follower arms positions
00:00
30K
Ville🤖
@VilleKuosmanen
Jul 28, 2025
the reward function powering the reliable robot is also getting better - 100 likes and I will open-source it 😇
00:00
05:33
Ville🤖
@VilleKuosmanen
Jul 28, 2025
Another uncut video of robots manipulating objects - this time it only missed a grasp at the very end. One step closer towards a general method of optimising robot reliability - still a long way to go though!
25K
Ville🤖
@VilleKuosmanen
Jul 29, 2024
I bought a robot 🤖
17K
Ville🤖
@VilleKuosmanen
Feb 2, 2025
A new AI model achieved my highest repeatability so far, at around 85% success rates. If the objects are at the centre and easily visible and reachable, this increases further
00:00
27K
Ville🤖
@VilleKuosmanen
Feb 15, 2025
New Chinese VLA paper - Dex-VLA Notes below 🧵
00:00
20K
Ville🤖
@VilleKuosmanen
Jan 7, 2025
first rollout of a fine tuned robotics foundation model. significantly undertrained (probably less than 1% training steps the original paper took), will continue training but there’s limits to how far a commercial GPU takes you
00:00
17K
Ville🤖
@VilleKuosmanen
Aug 7, 2025
Open-sourcing my reward model bolt-on for ACT! Using the code in the repo (which is @LeRobotHF compatible) you can train a simple reward model like the one in the demo with no changes needed to your existing LeRobot dataset & easy inference
00:00
22K