Konnex world trains robots with human feedback

Konnex World said on May 18 that it uses human feedback on videos of robot tasks to improve safety and behavioral quality in its robotics models, according to a post from the project’s developer account. The post described a training loop built on actual robot performance rather than fully synthetic data and showed examples involving household pickup tasks and grip adjustments in simulated trials. Konnex’s public documentation describes a broader system in which robot tasks are submitted to competing AI “miners,” scored by independent validators and recorded through what it calls Proof-of-Physical-Work. The company says its public testnet is already live with three workload classes — drone navigation, roboarm vision-language-action, and SLAM 3D mapping. ### What exactly did Konnex say it is doing? A May 18 post from Konnex World’s developer account said the project applies real human feedback to robot task videos to improve model safety and behavior quality. The examples in the post focused on manipulation tasks, including picking up household objects and adjusting grip during simulated runs. Konnex did not publish benchmark figures in the material reviewed, and the post framed the result as improved reliability in task outcomes rather than as a measured comparison against named baselines. The account’s examples were presented as simulated trials, not a disclosed commercial deployment. ### How does that fit with Konnex’s broader system? Konnex’s documentation says the network is designed around physical tasks rather than general chatbot outputs. Its docs describe a flow in which a task is signed and anchored onchain, multiple miners submit candidate policies or trajectories, and validators score execution on factors including safety, task match and efficiency. The same docs say the company’s “AI dashboard” evaluates competing policies across metrics such as accuracy, safety, speed, energy efficiency and stability before selecting a top result for execution on a real robot or a certified high-fidelity simulator. Konnex says verifiers then confirm execution evidence and post scoring records tied to on-chain payouts or penalties. (docs.konnex.world) ### Why use human feedback on robot videos instead of synthetic-only training? Konnex’s own materials argue that robotics needs verification tied to real execution records, not only model quality measured in abstract benchmarks. Its design overview says many current deployments still rely on operator-attested logs, while Konnex is trying to bind policy traces and sensor evidence into a record that outside validators can score. (docs.konnex.world) That framing helps explain the emphasis on human feedback over actual task videos. In Konnex’s setup, the useful training signal is not only whether a robot completed a task, but how it behaved while doing it — including grip, trajectory and safety-related choices visible in recorded execution. ### What is live now, and what is still staged? Konnex’s docs say the public testnet is live now on a Substrate-based layer-1 network with public RPC access, an explorer and three workload-class subnets. (docs.konnex.world) Those subnets are drone navigation, roboarm VLA and SLAM 3D map, according to the documentation. The roadmap language in the docs also says some pieces remain staged. Konnex says mesh transport, full robot-to-robot contract products and stablecoin-denominated mainnet settlement are not all live in the first public testnet release. ### What models and evaluation stack does Konnex say it uses? Konnex’s product documentation says it has onboarded three VLA starting points: OpenVLA, OpenVLA with OFT fine-tuning, and PI-0.5. (docs.konnex.world) The same page says a cloud-based 3D physical environment is used for safe rollouts and test iterations. The docs also say its verifier layer uses a vision language model to generate structured JSON metrics from task frames and instructions, and that the verifier stack is designed to be swappable or combined by consensus. (docs.konnex.world) Konnex’s next public reference points are its live testnet, builder documentation and workload subnets, all listed on the project’s docs site as of May 19. (docs.konnex.world) The company’s materials say dashboard URLs for some subnet views are tied to release notes as those interfaces are finalized. (docs.konnex.world)