sudo.ai develops sudo R1, an AI-driven robotic manipulation system trained exclusively on simulation data to achieve production-grade picking performance in real-world environments. The system delivers near-perfect success rates across diverse objects—including transparent, deformable, and reflective items—without any real-world demonstrations or fine-tuning. Its closed-loop control operates at 15–25 Hz, enabling real-time adaptation to dynamic scenes and obstacle-constrained spaces.
Funding
Funding not disclosed
Founders
Product
Problem
Robotic manipulation systems typically require extensive real-world data collection through teleoperation or manual labeling, which is slow, expensive, and insufficient to cover the full distribution of objects and conditions found in real environments. Existing systems often excel at individual capabilities like generalization or robustness but fail to achieve them simultaneously, limiting their production readiness.
Solution
sudo.ai's #sudo R1 is a robotic manipulation policy trained entirely on simulation data, eliminating the need for real-world demonstrations or manual labeling. The system achieves zero-shot generalization across diverse objects—including rigid, deformable, opaque, transparent, matte, and reflective items—with near-perfect robustness under changing lighting, dynamic backgrounds, and physical interference. #sudo R1 operates with true closed-loop agility, conditioning every action on the robot's latest observation at 15–25 Hz, enabling real-time adaptation to moving targets and mid-grasp perturbations. Its integrated spatial intelligence allows the policy to navigate around obstacles and through constrained spaces as a learned behavior, not a separate collision-avoidance module.
Target Audience
Primary customers are industrial automation providers, warehouse operators, and manufacturing companies seeking production-grade robotic picking solutions that generalize across diverse objects without per-object adaptation or extensive real-world data collection.
Features
- Trained exclusively on simulation data with no real-world demonstrations, teleoperation, or manual labeling
- Zero-shot generalization across unseen objects spanning rigid, deformable, opaque, transparent, matte, and reflective materials
- True closed-loop control at 15–25 Hz with every action conditioned on the latest observation, enabling reactive behavior to dynamic situations
- ~98% first-attempt success and nearly 100% within two attempts under changing lighting, dynamic backgrounds, random physical interference, and obstacle-constrained placements
- Integrated 3D obstacle awareness and viable-space reasoning for adaptive trajectory planning through constrained spaces
- Simulation-first scaling curve that improves capability by generating more data rather than scaling human labor