MambaAct speeds up π0.5 robots: action expert 10x smaller, latency down 64.8%

A team of researchers has presented MambaAct, a compact action expert built on the Mamba architecture for the π0.5 VLA model. It is 10x smaller than the original transformer expert and cuts flow-matching denoising latency by 64.8% in the LIBERO simulator and by 58.8% in real-world deployment. Quality did not drop; it rose. Average LIBERO success increased by 1.08 percentage points across three independent fine-tuning seeds. The code and weights have been released publicly.

Why the action expert became the bottleneck

VLA (vision-language-action) models take camera images and a text instruction and output a sequence of actions for a manipulator. In π0.5, two components handle this. A large vision-language prefix encoder interprets the scene and the command. A separate action expert generates continuous actions via flow matching: it starts from noise and "cleans" it into a trajectory over several iterations.

The problem is that word, "iterations." The encoder can run once, but the expert is called at every denoising step, and inside it the transformer repeatedly performs cross-attention over the prefix tokens. For a robot that has to react to the world in real time, these repeats translate directly into lost speed. This is where the MambaAct authors look for gains.

Mamba instead of a transformer, compressed context instead of cross-attention

There are two architectural changes. First, the transformer expert is replaced with a compact one in which Mamba works as an action mixer. It models dependencies within the noisy sequence of continuous actions. The pretrained π0.5 prefix encoder stays untouched, so its vision and language knowledge is preserved.

Second, without cross-attention, the expert still needs some way to learn what the robot sees and what it is being asked to do. For this, the authors added a lightweight Context Compressor. It turns the VLM prefix tokens into a compact conditioning signal that gates and modulates each Mamba block. Instead of attending to individual tokens at every step, the expert receives an already compressed representation of the context.

Key parameters of the work:

• base: π0.5 with a flow-matching action expert;
• action expert size: 10x smaller;
• denoising latency reduction: 64.8% on LIBERO, 58.8% in real-world deployment;
• average success gain on LIBERO: +1.08 percentage points across three fine-tuning seeds;
• evaluation on LIBERO-Plus and real-world tasks: the advantage holds;
• license and availability: code and weights published on GitHub.

Which matters more in the numbers: speed or quality

The 58.8% speedup on a real robot is more convincing than the 64.8% in simulation. Benchmarks forgive a lot, while a physical setup with cameras, actuators and delays shows whether the gain reaches the hardware. The gap between the two numbers is small, which is a good sign: the savings did not dissolve in real infrastructure.

The 1.08 percentage point gain in success is modest, but its value lies elsewhere. Shrinking an expert tenfold usually costs quality, yet here the metric went up. Confirmation across three independent seeds lowers the risk that this is a random spike. The advantage holding on LIBERO-Plus, where conditions are harder, adds another argument for the approach's robustness.

Where this is useful in practice

For teams already working with π0.5 or related flow-matching VLAs, MambaAct looks like a ready starting point for an experiment. There is no need to change the prefix encoder: the swap affects only the action expert, and since the weights and code are open, teams can compare results on their own tasks.

The biggest beneficiaries are those short on control frequency: mobile manipulators, platforms with weak onboard compute, and systems where a single GPU shares resources between perception and planning. A smaller expert frees up memory, and shorter denoising latency leaves more time headroom in each control cycle.

Caution is also warranted. The results come from LIBERO, LIBERO-Plus and specific real-world scenarios, and transfer to other robot morphologies and longer tasks will need separate testing. But the direction is clear: in robotics VLAs, it is worth optimizing not only the giant language component but also the small, repeatedly called action-generation block.