Publish a skill
A skill on Hyperspawn is a versioned, runnable physical-world capability. The same artifact ships to the browser sim, to AR previews, and to real Dropbears in the field. Skills carry a morphology envelope, not a hardcoded body: published policies are built to retarget to other humanoids, not only ours.
The fast path: hand it to your agent
Download the skill scaffold, fill in the brackets, and drop it into Claude Code or Codex in an empty repo. The agent sets up everything from that one file: motion retargeting from your demonstration videos, the simulated environment, reward and penalty terms, a training baseline, and the export artifacts below. Final calibration happens on the robot; optimizing the policy against reality is the brain's job, not yours.
- Record 1–5 short clips of the motion: you, a teleoperated robot, or an animation.
- Fill in
SKILL.md: demo, environment, rewards, penalties. - Tell your agent: "set this skill up." Train, evaluate, export.
- Upload the artifacts here.
What you upload
Exported from Isaac Lab / Newton. The runtime executes this in WebGPU.
IO shape, environment hints, robot compatibility. We validate it before publish.
Optional. Required for real-robot deploy. Contains the ROS2 node + launch files.
Contributor flow opens with the first batch
The upload UI is wired to the same artifact pipeline that the dropbear_isaac + dropbear_rl training repos export to. We're seeding it with the locomotion policies that already live in those repos before opening the form to the public.
Want to be in the first wave? Email hello@hyperspawn.co with a 30s clip of the policy and the repo it was trained in.
manifest.json schema (v1)
{
"name": "stable-walk",
"version": "0.1.0",
"category": "locomotion",
"robotModels": ["dropbear-v1"],
"io": {
"obs": { "shape": [42], "dtype": "float32" },
"action": { "shape": [22], "dtype": "float32", "bounds": [-1, 1] }
},
"control": {
"rate_hz": 50,
"actuator_map": "joint_names_v1.json"
},
"env": {
"trained_on": ["flat", "rough", "slope"],
"supported": ["sim", "ar", "real"]
},
"training": {
"framework": "isaac-lab",
"algorithm": "PPO",
"steps": 50000000,
"wall_hours": 14
}
}