Rei Labs

Adapt-1 Machina is a research release from Rei Labs that studies how continuous sequences of machine controls can be learned from coarse outcome feedback, without demonstrations or a critic. The release is aimed at researchers and robotics teams working on control policies, and it extends the earlier Adapt-1 Preview with experiments in sequence-based control across tasks like robotic manipulation, aiming across target configurations, and passing physical state between maneuvers. Each task defines its observations and available controls, and Adapt-1 proposes a sequence from the current observation, then uses execution feedback to revise later attempts. The work builds on Rei Labs' broader Adapt-1 program, which frames learning as something that happens at test time on a persistent substrate rather than through large-scale pretraining. Adapt-1 Preview was positioned as a pretraining-free substrate for test-time learning , and Machina is the applied follow-up that puts that substrate to work on physical control loops, including a YAM manipulation interface where the system proposes tool-pose and gripper commands that fixed inverse kinematics converts into joint targets. This launch matters now because it moves Adapt-1 from an architectural preview into concrete robotics experiments with reserved evaluation cases and defined learning budgets, giving outside researchers something specific to interrogate. Rei Labs describes itself as a research organization focused on advancing artificial intelligence through fundamental scientific principles rather than conventional computer science approaches , and Machina is the clearest public evidence yet of what that stance produces in a control setting, with API documentation available for teams that want to test the behavior directly.