We build in the open. Follow our engineering milestones as we develop autonomous litter collection robots — from simulation to deployment.
Our reinforcement learning policy reached 0.77 m/s sustained forward velocity on the 18-DOF quadruped model. The policy was trained using Isaac Gym with domain randomization across friction, payload, and terrain parameters. Per-joint clearance constraints keep feet above ground through the full gait cycle, and drift self-corrects without explicit heading control.
The full kinematic model of the CW-1 quadruped is defined in URDF with 18 degrees of freedom — 3 joints per leg (hip abduction, hip flexion, knee) plus 6 for the body-mounted gripper and bag system. The model is validated in RViz and used directly by the locomotion policy for sim-to-real transfer.
Our perception pipeline uses a custom-trained YOLO model that detects 50+ litter categories in real-time. The model runs client-side via WebGPU for our browser demos and on NVIDIA Jetson Orin for the production robot. Detection confidence averaging 0.85+ across common litter types including bottles, cans, wrappers, and cigarette butts.
The perception system combines stereo depth cameras, a forward-facing RGB camera for YOLO detection, and an IMU for terrain classification. Sensor fusion runs on the Jetson Orin with ROS 2 Humble, feeding a unified world model that the planner uses for obstacle avoidance, litter targeting, and path optimization.
Nine browser-based demos went live on cleanwalkerrobotics.com — covering litter detection, 3D robot visualization, quadrupedal locomotion physics, fleet management, route planning, and ROI calculation. All demos run client-side with no backend dependency, demonstrating our technology directly to prospects.
Low-level motor control firmware is being developed in Rust for safety-critical real-time control of the 12 leg actuators and 6 gripper/bag actuators. Rust's memory safety guarantees eliminate an entire class of firmware bugs common in C/C++ robotics stacks.
Schedule a call to get detailed technical briefings on our development timeline and pilot program availability.