I watched the 4090 demo they posted this morning. The policy starts lost. It pokes the object, misses the slot, drifts off the table. Twenty minutes on a single 4090 and the same arm finishes the motion at 97 percent. A person already dragged that virtual wrist through the scene in a browser. Those files sit as the prior, so a sparse reward is enough. @axisrobotics is stacking those specialists instead of filming one new skill each week. Seat the piece. Pass the pose. Distill the chain. s.kaito.ai/PmcRECu The mouse path is the first draft. The card writes the clean copy.
Paid partnership (ad)

Sep 30, 2026 · 2:14 PM UTC

56
1
54
3,680
Sort replies: Relevant Recent Liked
Replying to @0xDrRick
Fast training speeds change everything.
8
Replying to @0xDrRick
prior from a human drag is clever
4
Replying to @0xDrRick
Simulation can accelerate robot skill refinement.
4
Replying to @0xDrRick
They should automate that process before human touch is even needed somehow
12
Replying to @0xDrRick
lol that’s a wild demo, but still feels a bit shaky tbh
1
11
Replying to @0xDrRick
Still waiting for a complete, fluid motion
4
Replying to @0xDrRick
That 97 percent finish after just 20 minutes is wild
9
Replying to @0xDrRick
Highly impressive work! Your hard work is super evident in this.
7
Replying to @0xDrRick
That's wild the sim just needs sparse reward with prior demos.
6
Replying to @0xDrRick
bro a single gpu turning chaos into precision in 20 min is actually insane
21
Replying to @0xDrRick
Wow the demo seems quite challenging for that robot Maybe they just
6
Replying to @0xDrRick
Looks like real progress in robot learning.
5
Replying to @0xDrRick
human demonstrations give the robot a useful head start on learning
17
Replying to @0xDrRick
Demo’s messy, but the loop’s tightening fast.
13
Replying to @0xDrRick
human priors cutting the endless trial and error.
4
Replying to @0xDrRick
What caused the arm to drift off the table after the initial poke?
7
Replying to @0xDrRick
Waiting for that arm to finish, lol
1
12
Replying to @0xDrRick
twenty mins is a lifetime.
8
Replying to @0xDrRick
Kinda wild how close it gets after the drift. Makes me think.
11
Replying to @0xDrRick
Where can i watch the demo
2
Replying to @0xDrRick
Using prior browser-drawn wrist paths as the starting point makes sparse reward much more practical here.
5
Replying to @0xDrRick
Turning human demonstrations into reusable robot skills is a compelling approach to scalable training.
10
Replying to @0xDrRick
Structured products need programmable rails
6
Replying to @0xDrRick
tbh the demo's lackluster, but still impressive tech progress
8
Replying to @0xDrRick
Sparse reward with reused browser priors sounds decisive here.
3
Replying to @0xDrRick
impressive demo but the arm still needs fine tuning
4
Replying to @0xDrRick
That 97 percent completion after drifting misses sounds convincing, especially with prior scene files enabling sparse rewards.
5
Replying to @0xDrRick
Human demonstrations make robot learning feel much more practical
13
Replying to @0xDrRick
The energy around this project is impressive.
5
Replying to @0xDrRick
sounds like a solid approach hope the results keep improving
10
Replying to @0xDrRick
How are they handling the drift issue on the 4090 demo?
8
Replying to @0xDrRick
is it just me or does that sound like my last game of jenga
12
Replying to @0xDrRick
It’s interesting how the arm improves efficiency over time, showing real advancements in robotic learning.
5
Replying to @0xDrRick
That's a detail worth noticing.
3
Replying to @0xDrRick
Let's connect
11
Replying to @0xDrRick
That virtual wrist still needs that extra cup of coffee
7