Holo-M cuts the body of a humanoid into four action vocabularies and settles into the language model itself

Rather than a separate action expert, gestures become tokens added to the model's vocabulary. Grouped discrete broadcast decoding to hold real-time, and first places on the SIMPLE bench.

Discrete tokenized vision-language-action models work well to drive a robotic arm. A humanoid is a game-changer: the space of action of the entire body — legs, torso, arms, hands — is much larger and heterogeneous, which poses problems of tokenization, training and real-time inference that previous work did not address.

Holo-M is presented by its authors as the first discrete VLA model dedicated to humanoid locomotion-manipulation that intrinsically exploits the language model, by extending its vocabulary with action tokens. The tokenizer decomposes the action space into four specialized sub-tokenizers per body part — terminal effector, body, hand, kinematics — which makes it possible to train a mixture of very different sources: humanoid teleoperation, human video in egocentric view, simulation. By housing these tokens in the model's vocabulary, the authors avoid the isolation of knowledge induced by architectures with a separate continuous action expert.

For real-time, tokens from each body part are decoded by discrete mass diffusion instead of autoregressive generation. On the SIMPLE humanoid locomotion-manipulation testbed, Holo-M achieves the best success rates in generalist evaluation as a specialist, with a notable margin over the second. The code and weights are announced in open publication.