Hazard Intelligence

Data collection methods & models for humanoid mission tasks.

Hazard Intelligence records how people move on dangerous ground, rebuilds that exact ground in physics, and trains robots on it.

Ski slopes
40,000+
rescues on French ski slopes every winter
Mountain trails
2,890
search and rescue missions in US national parks in a single year
Backcountry
27
people killed by avalanches in an average US winter

Sources: French government winter prevention kit, 2023 · NPS incidents, 2017 · CAIC ten-year average

Data collection

Three ways in
Train from videosGoPro ski runs from YouTube, collected as links. We pull the body motion out and the robot learns from it.→ Train from gameplaySki in front of your webcam. Your motion capture becomes training data, and the robot copies you live.→ Train from your iPhonePhoton texts you when you head out. Say yes, and your phone records motion and GPS for that hike or ski day.→

Foundation models

Run it · one robot, zero-shot
Model API Pull a trained policy, architecture and weights included, and run it in your own physics loop at 50 Hz. No training on your side. Read the docs → 
from hi_client import pull

HOST = "api host"
policy = pull(HOST, "g1-ski")
actions = policy.infer(obs)   # 141 in, 31 out
Why robots fall

Robots doing complex tasks fail on the edge cases.

Physical Intelligence says it plainly: robots "can only operate in environments they've been trained in."

Robot data trails the data behind language models by an estimated 100,000 years, so labs now pay people by the hour to move in front of cameras. Almost none of it records the ground underfoot, so we tie every motion to the terrain it happened on and train robots on that same ground.

01 CapturePeople move; every sample is tied to the ground under it.
→
02 RebuildThat exact ground becomes physics.
→
03 TrainRobots practice there. Where they fall picks what to record next.
See how it works →
Terrain