Standardising how physical artificial intelligence is tested

Robotics engineering often stumbles during the transition from controlled laboratory settings to unpredictable real environments. Engineers frequently spend months setting up physical test tracks, conducting manual trials, and trying to diagnose why a trained model failed when lighting conditions changed slightly. Co-founders Raul Romero and Luigi D'Introno launched Kite ML to address this bottleneck in physical artificial intelligence. By building an evaluation layer and engineering harness, the company enables teams to augment training data, run multi-model trials, and evaluate policies prior to physical deployment.

Developed within the Founders, Inc. Canopy batch in San Francisco, Kite ML provides a dedicated environment for modern robotics workflows. Rather than relying solely on physical hardware runs, engineers can construct a digital twin of a workspace using just one teleoperation episode. Once created, teams can run thousands of simulated tests on a robot policy across different virtual environments at once. The system allows users to manipulate lighting conditions, swap objects, and introduce distractors to verify how robust a policy really is.

Integration is designed to fit directly into existing developer workflows without added friction. Teams can trigger evaluations through standard command line tools, direct API calls, or agentic coding tools like Claude Code and Codex via Model Context Protocol. Once a test run completes, the system outputs clear quantitative reports complete with statistical confidence intervals. This allows developers to monitor performance directly from their terminal or coding environment while keeping automated logs of every trial.

Crucially, the platform isolates exactly why a robot policy breaks down in practice. Instead of offering binary pass or fail results, the software automatically clusters failures by specific causes, such as object colour, clutter level, or improper grasp positioning. For instance, a policy might show high overall success but drop noticeably when encountering warm, dim lighting or multiple distracting items. This detailed breakdown highlights precise data gaps so engineering teams know exactly what information to collect for their next training iteration.

In addition to testing, the platform assists with data augmentation and multi-model training. Developers can evaluate multiple checkpoints and architectures simultaneously before deploying code onto actual hardware like Boston Dynamics Spot or WidowX arms. This approach helps policies generalise to unseen environments while avoiding costly physical crashes during early development phases. By reducing the time required to go from raw dataset to working policy, engineers can focus on complex tasks such as autonomous navigation and delicate object manipulation.

As Kite ML prepares to present at Web Summit 2026, the company focuses on advancing developer infrastructure for physical artificial intelligence. Raul Romero and Luigi D'Introno are focused on giving robotics teams the tools needed to move experimental models out of research labs and into functional real-world settings. Providing scalable simulation and continuous evaluation layer tools brings standard software rigor to physical hardware, helping teams build reliable robotic systems with higher confidence.

Find out more at kiteml.com