One Word, Different Action: A Real-Robot Benchmark for Language-Conditioned Embodied Reasoning

Robots that follow open-ended language instructions need to connect semantic intent to visual scene understanding, geometric feasibility, object states, and physical interaction conditions. End-to-end Vision-Language-Action policies have improved cross-task generalization, but they typically map visual and language inputs directly to robot actions, leaving limited explicit structure for long-horizon decomposition, physical verification, and recovery. We present \method, a zero-shot hierarchical framework functionally inspired by the division of roles in the human brain, comprising visual perception and state inference, language grounding and action-sequence generation from a shared atomic action library, cost-based plan selection, and real-robot execution and verification. The framework grounds commands in explicit object states, composes reusable atomic actions into task-conditioned sequences, ranks alternative sequences by execution cost, and verifies intermediate physical outcomes from refreshed observations. In the evaluation, \method{} completes 10/10 clean board trials, 10/10 pick-and-place trials, and 4/5 pyramid stacking trials for both the flat and irregular initial-layout conditions; the corresponding mean task progress is $99.03\%$, $100.00\%$, and $96.67\%$ respectively. Across all evaluated conditions, \method{} achieves higher success rates than ReKep, Dream2Flow, and $π_{0.5}$ benchmarks, demonstrating the effectiveness of combining explicit object-state reasoning, compositional atomic actions, cost-based plan selection, and closed-loop execution verification.

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/.

Robots following persistent language instructions must decide which cached room judgments remain valid as their environment changes. We present MIND (Maintenance via INstruction-conditioned Dependencies), a framework for maintaining target-room decisions without re-evaluating every room after each event. A language model compiles each instruction into dependencies for semantic qualification and target selection. The controller intersects semantic dependencies with event footprints, refreshes affected judgments, and updates reachability and selection separately. For fixed deterministic evaluation, planning, and selection, conservative dependency and footprint coverage preserves the decisions of full recomputation. On 240 five-event ProcTHOR sequences with an exact benchmark predicate, MIND reduces mean semantic evaluations from 30.6 to 11.2 while preserving all 1,200 post-event decisions. Frozen-evaluator experiments with Phi-4 and Qwen3-14B also preserve all 1,200 decisions, without improving the evaluators' underlying task accuracy. In 32 closed-loop AI2-THOR episodes using exact-predicate qualification and metadata-derived footprints, MIND matches the final target decision of full recomputation in every episode. Unitree G1 experiments demonstrate target reuse and reselection across room transitions without additional semantic evaluations under MIND, and abstention after a barrier and an object removal at 4 evaluations against 15 to 18. Additional materials are available at https://mind-icra27.github.io/.