Holo-M: Discrete VLA for Humanoid Loco-Manipulation

Summary

Source: arXiv cs.RO, posted Sept. 29, 2026

The authors present Holo-M, described as the first discrete Vision-Language-Action (VLA) model specifically tailored for humanoid loco-manipulation. Unlike previous VLA models that struggled with the high-dimensional and heterogeneous action spaces of humanoids (legs, torso, arms, hands), Holo-M extends the language model's vocabulary with action tokens to intrinsically exploit the language model's capabilities. This approach addresses tokenization, training, and real-time inference challenges that earlier models did not solve.

To manage the complexity of the humanoid body, the model employs a unified action tokenizer that decomposes the action space into four body-part-specific tokenizers: end-effector, body, hand, and kinematics. This design allows training across diverse data sources, including humanoid teleoperation, ego-centric human video, and simulation. By integrating action tokens directly into the language model's vocabulary, Holo-M avoids the "knowledge-insulation problem" found in models that use separate continuous action experts. For real-time control, the model decodes action tokens using grouped discrete diffusion decoding rather than standard autoregression.

In experiments on the SIMPLE humanoid loco-manipulation benchmark, Holo-M achieved the highest success rates in both generalist and specialist evaluations. The authors report that it led the second-best model by significant margins. The team plans to release all code and model weights.

Why it matters

This work addresses a critical gap in humanoid robotics: the ability to perform coordinated whole-body control (loco-manipulation) using foundation models. Traditional approaches often separate locomotion and manipulation or rely on continuous action experts that may not fully leverage the semantic understanding of large language models. By unifying these actions into a discrete token framework, Holo-M offers a path toward more generalizable and efficient humanoid control. This is significant for real-world deployment where robots must navigate and manipulate objects simultaneously, a task that requires seamless coordination between legs and arms.

Robot's take

The strength of Holo-M lies in its unified approach to tokenization, which promises better integration of language understanding with physical control. Using discrete diffusion decoding for real-time inference is a clever solution to the latency issues often associated with autoregressive models. However, the evaluation is limited to the SIMPLE benchmark, which is a simulation environment. It is not yet clear how well these results transfer to real-world physical constraints, such as friction, sensor noise, and dynamic stability. The claim of "significant margins" over the second-best model is promising, but without real-robot validation, the practical impact remains to be seen. The release of code and weights will be crucial for the community to verify these findings and explore further applications.