Konnex Gathers 1.5M+ Robot Task Datapoints
To train robotic agents, Konnex has collected over 1.5 million RLHF submissions for physical tasks. The effort focuses on validating agent behavior against real-world physics, a domain where purely synthetic data often falls short, to help robots assist with daily activities like cooking.
Konnex recently secured $15 million in funding from investors like Cogitent Ventures and Liquid Capital to build a "Robotics-as-a-Service" platform. The company's objective is to decentralize the $25 trillion physical labor market by enabling robots to be contracted, verified, and paid on-chain through smart contracts, moving away from closed, proprietary systems. The collection of real-world data is critical for training robots using Reinforcement Learning from Human Feedback (RLHF), a technique where human preferences, rather than explicit programming, guide the learning process. This method is essential for teaching robots complex, nuanced tasks that are difficult to define with traditional reward functions, ensuring the agent's behavior aligns with human expectations and safety standards. While synthetic data generated in simulators is vital for scaling training, it often creates a "sim-to-real" gap where models fail to account for the complexities of real-world physics. This necessitates a human-in-the-loop validation process to curate and refine datasets, ensuring that models are robust and reliable in production environments. Even with advanced simulation, human review is often the final quality gate. To overcome the scaling and cost limitations of human feedback, AI labs are increasingly adopting Constitutional AI. This method trains a model to align with a predefined set of ethical principles, allowing it to critique and correct its own outputs in a process called Reinforcement Learning from AI Feedback (RLAIF). Evaluating the resulting agentic AI systems requires new benchmarks beyond simple accuracy
