Perceptron AI ships ready-to-run 36B robot checkpoints, robotics gets its own pretrained economy

rohanpaul_ai · x · 2026-11-21

Perceptron AI released ready-to-run Hugging Face checkpoints of its 36B-parameter Isaac 0.5 robot model, each fine-tuned for a specific robot arm (e.g. bimanual YAM). Owners download the matching checkpoint and control the arm via plain-text instructions, with optional further fine-tuning. The bimanual YAM policy takes three RGB views, a language instruction and both arms' states, predicting absolute joint and gripper targets; it was fine-tuned to step 200,000 on 32 GPUs with a global batch of 256. Built on LeRobot, with code and datasets open-sourced.

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

Original post →

More from Embodied

Embodied channel →