Launching today
Kineforge AI/ML model for robotics
Train robot brains without demos on one GPU
5 followers
Train robot brains without demos on one GPU
5 followers
Kineforge trains embodied control policies without human teleoperation demos. Frozen semantic priors + differentiable physics (MuJoCo MJX) + a 157K-parameter policy = training in hours on a single GPU, and CPU inference under 10ms. Structure over scale.
Pixlie
Hey Product Hunt š I'm Illia, solo founder of Brightforge.
After Pixlie (our AI video studio), this is our research track: Kineforge.
The problem: teaching robots to act usually means one of two expensive paths ā
record thousands of human teleoperation demos, or run billion-parameter
"robot brain" models that only live in the cloud. Both are slow, costly, and
hard to deploy where it actually matters (warehouses, labs, hazardous sites).
Kineforge takes a different bet: structure over scale.
ā¢ No human demos. We train in simulation (MuJoCo MJX), not from teleop recordings.
ā¢ Tiny trainable policy ā 157K parameters (not billions), sitting on top of a
frozen "knowledge layer" that gives the agent structure from day one.
ā¢ Trains in hours on a single GPU, not weeks on a cluster.
ā¢ Runs on a CPU in under 10ms ā edge-deployable, no datacenter required.
On NVIDIA H200 we log stable task success ~0.70 with open metrics and honest
simulator artifacts (345K+ logged steps, 0 crashes on a 1.5M-step run).
This is early and sim-first, real-robot pilots are next. I'd love feedback from
robotics, RL, and embodied-AI folks, and intros to warehouse / lab automation
teams who feel the data-cost pain.
Imagine teaching a robot to do a job.
Today there are two slow, expensive ways:
1) Sit a human down for months to "puppet" the robot thousands of times
(teleoperation). Slow. Costly. Boring. Doesn't scale.
2) Rent a giant AI "brain" with billions of settings that only runs in a
massive cloud datacenter. Expensive every single time it thinks.
Kineforge does it differently. We give the robot a small, smart "starter brain"
that already understands structure ā like a kid who starts school already
knowing the alphabet, instead of inventing language from zero.
Then it practices inside a fast video-game-like simulator (no real-world risk,
no humans puppeting it). It learns the job in hours, on one computer chip ā
not weeks on a server farm.
The payoff for a business:
ā¢ 10ā50Ć less training data needed
ā¢ 5ā10Ć cheaper to develop each new skill
ā¢ The finished "brain" is tiny ā it runs on a cheap chip inside the machine,
in the blink of an eye, with no cloud bill per move.
Imagine teaching a robot to do a job.
Today there are two slow, expensive ways:
1) Sit a human down for months to "puppet" the robot thousands of times
(teleoperation). Slow. Costly. Boring. Doesn't scale.
2) Rent a giant AI "brain" with billions of settings that only runs in a
massive cloud datacenter. Expensive every single time it thinks.
Kineforge does it differently. We give the robot a small, smart "starter brain"
that already understands structure ā like a kid who starts school already
knowing the alphabet, instead of inventing language from zero.
Then it practices inside a fast video-game-like simulator (no real-world risk,
no humans puppeting it). It learns the job in hours, on one computer chip ā
not weeks on a server farm.
The payoff for a business:
ā¢ 10ā50Ć less training data needed
ā¢ 5ā10Ć cheaper to develop each new skill
ā¢ The finished "brain" is tiny ā it runs on a cheap chip inside the machine,
in the blink of an eye, with no cloud bill per move.
So instead of "AI that's smart but ruinously expensive to run," you get
"AI that's good enough AND cheap enough to actually put in thousands of
machines." That's the difference between a demo and a business.
We've proven the economics in simulation. The next step is putting it into a
real warehouse robot.
Under the hood:
ā¢ POMDP control, actions factored into locomotion (16) / respiration (2) /
head (2) / macro (4) = 24 discrete decisions per step.
ā¢ Multi-stream policy: Fast (reflex), Valence (affect-modulated gains),
Slow (topology-routed planning).
ā¢ Frozen semantic prior (~2.6M params, 110 canonical symbols, 768-dim) supplies
structured inductive bias; only ~157K policy params are trained (REINFORCE/PPO,
JEPA-style + structural regularization).
ā¢ Backend: MuJoCo MJX differentiable physics, JAX-native (no PyTorch in core),
SkyPilot on Kubernetes, NVIDIA H200, MLflow tracking.
ā¢ Results: ~0.70 task reward at 100K steps; ~159 (baseline) vs ~108
(grounded sensory) env-steps/s; CPU inference target <10ms.
ā¢ Sim-only today; sim-to-real via Unity/ROS2 in progress. Priors withheld;
sanitized eval protocol published.
So instead of "AI that's smart but ruinously expensive to run," you get
"AI that's good enough AND cheap enough to actually put in thousands of
machines." That's the difference between a demo and a business.
We've proven the economics in simulation. The next step is putting it into a
real warehouse robot.
Under the hood:
ā¢ POMDP control, actions factored into locomotion (16) / respiration (2) /
head (2) / macro (4) = 24 discrete decisions per step.
ā¢ Multi-stream policy: Fast (reflex), Valence (affect-modulated gains),
Slow (topology-routed planning).
ā¢ Frozen semantic prior (~2.6M params, 110 canonical symbols, 768-dim) supplies
structured inductive bias; only ~157K policy params are trained (REINFORCE/PPO,
JEPA-style + structural regularization).
ā¢ Backend: MuJoCo MJX differentiable physics, JAX-native (no PyTorch in core),
SkyPilot on Kubernetes, NVIDIA H200, MLflow tracking.
ā¢ Results: ~0.70 task reward at 100K steps; ~159 (baseline) vs ~108
(grounded sensory) env-steps/s; CPU inference target <10ms.
ā¢ Sim-only today; sim-to-real via Unity/ROS2 in progress. Priors withheld;
sanitized eval protocol published.
Model card and tech report is still wait endorsement on arXiv, because my CS teacher have only 2 endorsement by himself and to aprove you must have at least 3,sooo... (no fluff, real tables):
https://illiaovcharenko1.github....
Links
PDF (v1.0): Tech report PDF
Landing: GitHub Pages
GitHub: kineforge-portfolio
HuggingFace (HF Space): kineforge-demo
arXiv: pending endorsement (cs.RO) ā then https://arxiv.org/abs/ā¦
Contact: illia.ovcharenko@brightforge.ltd Ā· brightforge.ltd
Try the demo + ask me anything below š