Technical Report v1.0 · Brightforge Software Inc.

Kineforge

Sample-efficient embodied control with frozen semantic priors (~2.6M fixed) and a compact trainable policy (~157K) on differentiable physics (MuJoCo MJX).

Download PDF (v1.0)

Tech report

PDF from Overleaf. arXiv ID will be linked here after endorsement + submit.

PDF · GitHub (TeX + figures)

arXiv: pending endorsement (cs.RO)

Live demo

Gradio mock — no weights in public repo.

Hugging Face Space

Contact

Brightforge Software Inc. · Kineforge AI-Lab.

illia.ovcharenko@brightforge.ltd

brightforge.ltd

Metrics

Nebius H200 — ~159 vs ~108 env-steps/s; task reward ~0.70.

Learning curve (MLflow)

Homeostasis vs training steps — monolithic run.

Homeostasis learning curve

Throughput vs task reward

Throughput scatter