This paper introduces a simple post-training recipe that turns off-the-shelf VLMs into robotic reasoners, and suggests that free-form language reasoning can function as a test-time compute mechanism for steering low-level policies.
Abstract
Reasoning in language allows foundation models to spend more test-time compute on hard problems, such as those requiring decomposition, constraint tracking, and prediction of future consequences. Whether this mechanism can improve robotic manipulation remains unclear, where long-horizon tasks require tracking partial progress, reasoning about object relations, recovering from mistakes, and steering noisy low-level policies. In this paper, we study whether VLMs can be trained to reason directly in natural language to guide low-level manipulation policies. We introduce $R^3$, a simple post-training recipe that turns off-the-shelf VLMs into robotic reasoners: it first mid-trains a VLM on expert-generated reasoning traces to initialize the desired reasoning style, then improves the reasoner with single-step rubric-based RL from offline action data. Unlike prior robotic reasoning methods that mostly use structured traces as auxiliary supervision, $R^3$ trains free-form language reasoning to produce test-time guidance for action. We instantiate $R^3$ on Language Table and simulated bimanual grocery packing, two controlled testbeds for studying robotic reasoning and long-horizon manipulation. $R^3$ improves exploration and generalization across unseen tasks and significantly outperforms instruction-only imitation learning baselines on both benchmarks. Our analyses suggest that free-form language reasoning can function as a test-time compute mechanism for steering low-level policies. Our project page is available at https://robotic-reasoner.github.io/.
Continuous robotic control requires policies that execute with low latency and modest computational cost during deployment. Foundation models provide strong semantic and visual reasoning, but repeatedly querying a large model throughout deployment incurs substantial inference latency and compute requirements. Language-to-Reward (L2R) methods avoid this deployment-time cost by using large language models (LLMs) to synthesize rewards for training lightweight policies, but these rewards are generated without visually analyzing how the learned policy physically fails, and thus often lack physical grounding. We propose Visually-Grounded Reward Synthesis (VGRS), which uses slow foundation models during training to produce fast robotic control policies. An LLM first synthesizes executable reward code from a natural-language instruction to train a lightweight hierarchical policy. When learning stalls, a frozen vision-language model (VLM) analyzes failed trajectories to provide failure mode diagnosis, which the LLM uses to rewrite and densify the reward. Since foundation models are used only during training, deployment requires only the learned policy. We perform experiments on simulated and real-world navigation and manipulation tasks, and show that VGRS achieves success rates above 55% on challenging long-horizon tasks while deploying successfully to real robots.
Utsav Singh, Pramit Bhattacharyya, Vinay P. Namboodiri· 3 citations
Vision-language-action (VLA) models are trained by imitation and capture what action to take but not why; adding causal reasoning improves manipulation, but current methods pay for it at inference time - generating reasoning tokens or rolling out predicted future states at every step, a cost that compounds over long horizons. We ask whether this benefit can instead be captured during training and discarded before deployment. We introduce Latent Semantic Scaffolding (LSS), an auxiliary loss applied during human-demonstration pretraining that aligns a VLA's action-token representations to text embeddings of physical-reasoning rationales through a small projection head. The head is dropped at inference, leaving the unmodified base policy with zero added cost. Our central finding concerns alignment granularity: aligning each action token to the rationale of its own manipulation phase (Dense LSS) rather than to a single pooled episode-level embedding (Pooled LSS) yields representations that transfer markedly better to held-out tasks. Dense LSS attains both the best in-distribution success and the best transfer to tasks unseen during alignment, whereas pooled alignment over-specializes to the training task. A representational probe shows Dense LSS induces roughly twice the per-phase separability in the backbone, supporting that phase-local alignment is the operative mechanism.
Across in-domain and distribution-shifted settings, allocating additional test-time computation substantially improves next-subtask prediction accuracy, and these gains translate into higher closed-loop success on long-horizon robot manipulation tasks.
Xiaowei Cai, Yunuo Cai, Bing Chen et al.· 0 citations
The approach, LeAct (Learning to reason from Actions), optimizes this latent variable: the student samples candidate CoTs for each expert action, and the student retains those that measurably improve its own probability of recovering the action.
Hierarchical Robotic Control (HiRoC) is proposed, a hierarchical post-training framework that decouples high-level task planning from low-level action execution and aligns the executor with planner-generated subgoals before reinforcement learning, mitigating the distribution misalignment between planning and execution.
Two systematic attempts to improve large pretrained models with minimal or zero modification to their weights via reinforcement learning on a frozen OpenVLA-7B using binary task-success rewards on LIBERO-Goal reveal a common ceiling.
Adam Lalani, Chen Sun, Hui Wang· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
