Perceptron Releases Ready-to-Run Checkpoints of Its 36B Robot Model Isaac 0.5

rohanpaul_ai · x · 2026-11-21

Robotics is getting the pretrained-checkpoint economy that sped up language models: Perceptron AI released ready-to-run versions of its 36B-parameter robot model Isaac 0.5 on Hugging Face, each fine-tuned for a specific robot arm.

Owners of supported arms (YAM, SO-100/101, LIBERO) can download the matching checkpoint and have the arm act on plain-text instructions, with optional further fine-tuning. The YAM bimanual checkpoint was fine-tuned on xDoF ABC and MolmoAct2 demonstrations, reading three RGB views, a language instruction, and both arms' states to predict absolute joint and gripper targets — saved at step 200,000 on 32 GPUs and supported through LeRobot.

Related event: Perceptron releases ready-to-use 36B Isaac 0.5 robotics checkpoints(2 posts)→

Original post →

More from Embodied

Embodied channel →