Zeroth
Zeroth offers the M1, a 50 cm desktop-sized humanoid robot designed as an accessible entry point for reinforcement learning research. The 12-DOF bipedal platform mirrors full-size humanoid leg structure, enabling researchers to train and deploy RL policies that transfer directly to human-scale robots. It supports PPO algorithms, NVIDIA Isaac Gym and MuJoCo simulation, and includes pre-loaded reference behaviors like walking, obstacle traversal, and getting up.
- Robotics
- Artificial Intelligence
- Consumer Electronics
- Hardware
- Commerce & Shopping
Funding
Founders
Founder details are not available yet.
Product
Problem
Researchers and developers seeking to work on humanoid robotics often face significant barriers, including the high cost and risk of full-size humanoid platforms, limited access to shared lab equipment, and the need to start with quadruped robots as a detour before advancing to humanoid systems. This makes it difficult to build and test reinforcement learning algorithms on humanoid platforms in a practical, accessible way.
Solution
Zeroth provides the M1, a 50 cm desktop-sized humanoid robot designed as an entry point for reinforcement learning on humanoid platforms. The M1 features a 12-DOF bipedal leg structure that matches full-size humanoid configurations, allowing users to train RL policies that transfer directly to human-scale robots. It runs on industry-standard PPO algorithms and supports NVIDIA Isaac Gym and MuJoCo simulation platforms, with a Python API for training and deployment. The robot includes onboard sensors such as LiDAR, RGB camera, iTOF depth sensor, and microphone array, along with pre-loaded reference behaviors including walking, obstacle traversal, getting up, and dancing. The company also offers the W1 outdoor explorer robot and Jupiter humanoid robot for broader applications.
Target Audience
Primary customers are robotics researchers, AI developers, and engineering students who want to build reinforcement learning skills on a humanoid platform without the cost and complexity of full-size robots, as well as technology enthusiasts interested in consumer-grade humanoid robots.
Features
- 12-DOF bipedal leg configuration (6 DOF per leg) matching full-size humanoid structure for transferable RL research
- PPO reinforcement learning support with pre-tuned walking reward functions and customizable reward templates
- Sim-to-real pipeline compatible with NVIDIA Isaac Gym and MuJoCo, with PyTorch/ONNX model export
- Python SDK and M1 Spine SDK for low-latency control, with ROS 2 integration and Linux + OTA workflows
- Onboard perception stack including head LiDAR, RGB camera, iTOF depth sensor, and 3-microphone array
- Six pre-loaded reference behaviors: variable speed walk, obstacle traversal, one-leg kick, get up, dance, and moving platform stabilization
- One-line terminal deployment command or controller shortcut (LB + X) for policy deployment
- Customizable face display interface for expressions, moods, and status indicators