HALO: A Unified Vision-Language-Action Model for Embodied Multimodal Chain-of-Thought Reasoning

Vision–Language–Action (VLA) models perform well in robotic manipulation, but often struggle in long-horizon or out-of-distribution scenarios due to the lack of explicit mechanisms for multimodal reasoning and anticipating how the world evolves under action. Recent works introduce textual chain-of-thought or visual subgoal prediction within VLA models, but still fail to offer a unified human-like framework for joint textual reasoning, visual foresight, and action prediction. To this end, we propose HALO, a unified VLA model that enables embodied multimodal chain-of-thought (EM-CoT) reasoning through textual reasoning, fine-grained visual subgoal prediction, and EM-CoT-augmented action prediction. We instantiate HALO with a Mixture-of-Transformers (MoT) architecture that decouples semantic reasoning, visual foresight, and action prediction into specialized experts with seamless cross-expert collaboration. To enable HALO learning at scale, we introduce an automated pipeline to synthesize EM-CoT training data along with a carefully crafted training recipe. Extensive experiments demonstrate that: (1) HALO achieves superior performance in both simulated and real world, surpassing baseline policy $\pi_0$ by 25.9\% on the RoboTwin benchmark; (2) all proposed components of the training recipe and EM-CoT design help improve task success rate; and (3) HALO exhibits strong generalization under aggressive unseen environment randomization with our proposed EM-CoT reasoning.

Most robotic systems today map raw visual inputs directly to motor commands, which tends to break down when facing novel objects, unfamiliar layouts, or tasks requiring sustained reasoning across many steps. We introduce HALO, a unified VLA model that replicates the human cognitive sequence of thinking, imagining, and then acting. Before generating any physical movement, HALO first produces a textual reasoning trace articulating the current subtask and its role in the broader plan, then generates a visual prediction of what the scene should look like once that subtask is complete, and only then commits to a concrete sequence of actions. This is realized through a mixture-of-experts architecture that assigns language understanding, image generation, and action control to three specialized modules sharing a common attention mechanism, alongside an automated pipeline that annotates raw robot demonstrations with reasoning traces and goal images at scale. Experiments in both simulated and real-world environments show that grounding actions in explicit multimodal reasoning leads to substantially more robust behavior, particularly in out-of-distribution scenarios involving unfamiliar objects, lighting changes, or visual distractions.