Introducing
DROPBEAR
A humanoid robot that walks, learns, and does what you ask. We have a platform for you to start building your humanoid robot today. Or buy one.
It learned to walk.
The gait is a learned policy, not a scripted trajectory. Dropbear trains by reinforcement in NVIDIA Isaac Lab, whole body and per subassembly, across thousands of parallel simulations, and transfers to hardware. Onboard, the balance loop closes at 500 Hz. The toggle above is the same policy at each stage: training, simulation, camera.
Built for human spaces.
The world is sized for the human body: stair treads, door handles, shelf heights, tool grips. A humanoid fits all of it without changing any of it. Thirty-two joints, five-fingered hands, five kilograms of payload per arm, repeatable to a tenth of a degree.
Ask. It does.
Dropbear runs on language. Say what you need and it works out the steps, picks the skill, and picks up the tool. Watch it take requests.
Two ways in.
Same robot at the end of both roads. One arrives in a crate; one comes off your printer.
It learns new tricks overnight.
Skills are versioned policies: trained in simulation, installed over the air, calibrated on the robot. The brain watches the situation and pulls what the moment needs. And skills retarget across bodies, so what one robot learns is built to run on every humanoid. Not just ours.
Locate, grasp, set down. The primitive behind every chore.
Pick it up and walk it anywhere. Five kilograms, either arm.
Tool use: grip, align, drive. Human tools, unmodified.
The kitchen is a workspace like any other.
Doors, latches, buttons. The small stuff is the hard stuff.
Video in, policy out. Publish a skill and every robot can learn it.
Build the
humanoid era.
We are heading into a world where robots follow instructions and useful work stops being scarce. Every file here is public. Every robot is owned, not rented. What that future looks like is up to the people who build it.