Tomorrow Robotics unveils Habilis Brain 0 foundation model, reporting 99.55% on LIBERO and 400 straight picks on Rainbow's RB-Y1

Seoul physical-AI startup Tomorrow Robotics, co-led by Zhang Byoung-tak and Park Sang-wook, said on Oct 7 that it has released Habilis Brain 0, a robot foundation model for grasping and moving objects, Robot Newspaper (irobotnews) and Edaily reported; MEKAWIRE did not find the company's own release, so the benchmark figures are company claims as reported.

According to the reports, the model pairs GC-VLA, a vision-language-action model that learns how the hand and object positions should change during a task, with GCRF, a reinforcement-learning step that corrects the base policy from success and failure experience; the company said GC-VLA alone scored 95.20% on the LIBERO simulation benchmark and the combined model 99.55%.

On hardware, it said the model completed 400 consecutive logistics picks on Rainbow Robotics' dual-arm mobile RB-Y1 after starting from about three hours of demonstrations plus robot-collected and human-corrected data, ran a two-handed container-packing task on a Universal Robots UR5e and recovered from dropped or moved objects on Robotis' AI Worker.

Tomorrow Robotics said it is commercialising the model with Rainbow Robotics and Robotis under existing MOUs and has been selected for the SME ministry's Scale-up TIPS, AWS's Physical AI Frontier and Intel's Ingenious programmes.

MEKAWIRE has no other 2026 story with a Korean company headlining a robot foundation model, a field so far led in our coverage by U.S. and Chinese labs, and running one model across three makers' arms is the claim worth testing.