VLA
Vision-Language-Action model

VLA models integrate visual input, language input, and action output into a single learning or planning framework so that grounded semantics from language map to perceived affordances and actionable policies; architectures typically combine joint or aligned vision-language encoders (often transformer-based or cross-modal attention mechanisms) with policy modules trained via imitation learning, reinforcement learning (RL), or differentiable planners to produce temporally-extended, sensorimotor behaviors. Their significance lies in grounding symbolic instructions in continuous perception and control — enabling tasks such as instruction-following navigation, object manipulation, and interactive dialogue with embodied agents — and in advancing research on generalization, sample efficiency, and safety in embodied AI by bridging representation learning (vision-language alignment) with control theory and sequential decision making.

First used in the context of Vision-and-Language Navigation (VLN) research around 2018 (e.g., the Room-to-Room work), the explicit framing of "vision–language–action" became more widely used and influential between 2020–2022 as benchmarks (ALFRED, Habitat) and advances in large-scale vision-language models and policy learning pushed embodied and multimodal agent research into broader prominence.

Key contributors include research groups and communities bridging multimodal ML and robotics: teams behind VLN and R2R (e.g., FAIR/Meta and collaborators such as Peter Anderson) and the ALFRED benchmark (Shridhar et al.) that formalized instruction-to-action tasks; simulator and dataset creators (Habitat, Matterport3D, Gibson) that enabled large-scale training and evaluation; and institutional contributors advancing the underlying components — OpenAI and others for scalable vision–language pretraining (e.g., CLIP-like models), DeepMind and robotics labs for policy and RL advances, and university labs (Berkeley, Stanford, CMU, AI2) for foundational work on grounded language, imitation learning, and sim-to-real transfer.