Kineforge — Portfolio One-Pager

<aside> ⚙️

Kineforge trains physical agents in differentiable physics (MuJoCo MJX) with a compact multi-stream policy guided by frozen structured semantics — hours on one GPU, <10ms CPU inference.

</aside>

At a glance

Training: JAX + MJX rollouts
Deployment: CPU edge
Focus: sample-efficient robot learning (no human demos)

What you get

Faster iteration vs. teleop-heavy pipelines
Better coverage in hazardous/remote environments
A path to sim-to-real via Unity / ROS2

One-liner

Kineforge trains physical agents using differentiable physics (MuJoCo MJX) and a 157K-parameter multi-stream policy guided by 2.6M frozen semantic priors — hours on one GPU, <10ms CPU inference.

Problem

Robotics today needs massive teleoperation datasets or billion-parameter cloud models
This breaks in toxic, remote, or expensive environments where humans can’t demo

Solution

Layer	What it does
Frozen semantic prior	Structured inductive bias (110 symbols, not trained in RL)
Fast stream	Reflexive locomotion + head (emotion-modulated)
Valence stream	Affective state → exploration stability
Slow stream	Topology-routed macro planning (22-D route manifold)
MJX physics	High-throughput differentiable rollouts

Traction

Built

✅ Working JAX training pipeline on Nebius H200
✅ MLflow + SkyPilot + S3 checkpoints

Measured

✅ 345K+ environment steps (monolithic jobs); 100K A/B studies
✅ Task reward ~0.70, baseline ~159 steps/s

In progress

🔄 Sim-to-real: WebSocket → Unity / ROS2
👥 1 industrial pilot (warehouse), 3 academic alpha users

One-liner

Problem

Solution

Traction

Why us vs. others