Armen Aghajanyan cofounded Perceptron with Akshat Shrivastava after leading multimodal research at FAIR. His work connects mixed image-text modeling with systems that perceive and act in physical environments.[1][2][3]
Start with From VLM/VLAs to Embodied Agents →
Links
Perceptron
X Cofounder and CEO · Perceptron
Career
Aghajanyan and Shrivastava started Perceptron in November 2024 after leading multimodal research initiatives at FAIR. The company develops models that process video, audio, text and sensor data, and supports customers from integration through production. Its model program includes the open Isaac family and closed models for image, video and embodied reasoning. The work centers on interpreting changing physical environments, rather than only producing answers about static text.[1]
Contributions
Generate across a mixed document
Aghajanyan is first author of CM3. It trains a generative model on structured documents containing text and image tokens. A causal-masking objective removes spans from their original positions and generates them at the end of the sequence. That preserves left-to-right generation while enabling the model to complete missing material using surrounding context. The paper examines both language tasks and multimodal generation.[2]
Early fusion for images and language
Chameleon treats image and text tokens as parts of one sequence from the start of training. The model can understand and generate interleaved image-text documents rather than connecting separate specialist models only at a late stage. The report discusses stability, alignment and architectural choices needed for this setup. Its contributor statement identifies Aghajanyan as a joint first author and pretraining workstream lead, distinguishing that role from overall project leadership.[3]
Ideas
In his Perceptron talk, Aghajanyan examines how vision-language and vision-language-action models can become parts of embodied agents. The proposed system has to select useful information from continuous visual input and connect perception to reasoning and control. His account frames efficient processing of video as a constraint on practical physical intelligence.[4]
Tools, writing & talks
Writing, research & talks
Start here
From VLM/VLAs to Embodied Agentsai.engineerA talk connecting multimodal perception with systems that act in the physical world.
Start with the examples that connect a visual observation to an action, then follow the discussion of processing video efficiently. Familiarity with vision-language models helps, but no robotics implementation is required.
Speaker[4]Perceptron AI's Armen Aghajanyan explains a perceptive training objective and sparse mixture-of-experts routing built to avoid wasting compute on video's mostly uninformative pixels, underlying a model he claims beats a frontier lab's embodied reasoning model at far lower cost.
Catalog workOriginal sourceA generative model and training objective for structured documents containing text and images.
First author[2]Explains training a unified model for interleaved image and text sequences.
Joint first author and pretraining workstream lead[3]