Tommoro Robotics has introduced Habilis-β, an on-device robot AI model developed for real-world industrial applications.
Habilis-β was evaluated on productivity and reliability during sustained operation, going beyond conventional evaluations focused on success in individual attempts. Instead of asking whether a task could be completed once, the evaluation measured how quickly and reliably the robot could repeat tasks continuously for one hour under conditions closer to industrial use.
In the RoboTwin 2.0 simulation environment, the model continuously performed manipulation tasks including lifting and pouring objects into a large container, positioning a pair of shoes, and stacking three bowls in sequence. It achieved 572.6 tasks per hour (TPH), a substantial improvement over the comparison models in this evaluation.
Real-world testing used an RB-Y1 humanoid robot for Dual-Bin Conveyor Packing. Over a continuous one-hour run, the robot picked objects from bins on both sides and placed them into boxes on a moving conveyor. Habilis-β achieved 124 tasks per hour. Mean time between interventions (MTBI) also increased substantially, indicating improved reliability during continuous operation.
In this real-world evaluation, throughput improved by up to approximately six times compared with NVIDIA’s GR00T N1.5 and Physical Intelligence’s pi0.5. The results assess practical productivity and sustained-operation reliability, alongside task success.
Habilis-β automatically adjusts motion speed, distinguishing between phases that allow fast movement and those that require precision. Its lightweight computing architecture runs directly on the robot’s onboard computer. This reduces network dependence and latency to support the real-time responsiveness and operational reliability required in industrial settings.