Research

The substrate beneath
every Hyperspawn skill.

We train policies in Isaac Lab and NVIDIA Newton, package them with their ROS2 runtime, and ship them to the marketplace. Every skill on Hyperspawn is reproducible from the repos below — and runnable in your browser the moment it's published.

Browse trained skills Run an eval in browser

The training → deploy pipeline

step 1

Train

Isaac Lab / Newton, local + cluster. PPO, SAC, IL.

step 2

Export

ONNX policy + manifest.json + ROS2 bundle.

step 3

Publish

Versioned artifact in the Skills marketplace.

step 4

Deploy

Browser sim, WebAR, real Dropbears — same artifact.

Open repos

Isaac Lab

dropbear_isaac

Locomotion + manipulation tasks for Dropbear in Isaac Lab. PPO baselines, terrain curriculum, and the eval harness our /sim runs.

Newton

dropbear_rl

Contact-rich manipulation policies trained in NVIDIA Newton. The bridge layer that makes a single policy shippable to sim, AR, and real.

Hardware

dropbear-feet

TPU 95A printable feet with 4× 50kg load cells and HX711 80Hz mod. The proprioceptive layer beneath the locomotion stack.

Hardware

aero-hand-open

Open dexterous hand from TetherIA. Ships with grasp synthesis policies trained to run on day-one Dropbears.

Currently training

Live training feed lights up shortly.

When a contributor pushes a training run from Isaac Lab, it shows up here with live reward curves and an auto-published browser-runnable URL once it converges.

Reproduce a skill end-to-end

# 1. clone + install
git clone https://github.com/Hyperspawn/dropbear_isaac
cd dropbear_isaac && uv sync

# 2. train (single GPU, ~14h to convergence)
python -m dropbear_isaac.train --task walk --algo ppo --steps 50_000_000

# 3. export to ONNX + manifest
python -m dropbear_isaac.export --checkpoint runs/walk/best.pt

# 4. publish to Hyperspawn
hyperspawn skills publish ./out/walk/

# 5. it's now live at https://hyperspawn.co/sim?skill=walk