Excited to share LWD: Learning While Deploying. Our robots learn while doing real tasks—restocking groceries, brewing Gongfu tea, making cocktails, making juice, and packing shoes. Deployment is no longer just evaluation; it becomes the training loop. 🧵

Starting from a pretrained VLA policy, LWD deploys one shared generalist policy across a fleet of 16 real dual-arm robots. The fleet collects autonomous rollouts, failures, recoveries, partial progress, and human interventions from the real world.

This is the LWD data flywheel: robots deploy, collect real interaction data, update the shared policy with offline-to-online RL, and redeploy the improved model back to the fleet. Deployment becomes a continuous source of learning.

Technically, LWD combines distributional implicit value learning from heterogeneous fleet data and adjoint matching for policy extraction in flow-based VLAs. Across 8 real-world tasks, one generalist policy reaches 95% average success.

Website: Paper:

Cool!

Can we bring your robots to the biggest physical AI buildathon in Southeast Asia this July? 2K AI builders await!

this is the loop i keep coming back to. once robots are learning from deployment, you need open tools for all the unsexy stuff around it: serving the policy, logging rollouts, replaying failures, comparing versions, and not losing your mind every time the model changes. i’m building an open source tool called Reflex and looking for contributors + robotics people in the discord:

i dont understand why made in human format, just elevate cost, just a box with some "arm" is enough.

cool demo but without robust safety constraints learning while deploying bites hard

On s’en fout.