Bingqi Huang, Bingchuan Wei, Yingkai Cai +1cs.RO cs.CV
World action models (WAMs) jointly denoise future video frames and robot actions, and the video prior is expected to generalize their control. Camera viewpoint change remains one of their hardest perturbation axes. We study a question specific to this model class: when training with same-state cross-view image pairs, on which output coordinates should a consistency loss be imposed? The WAM denoising target mixes view-covariant coordinates, namely the predicted future scene, with view-invariant coordinates, namely the action chunk, future proprioception, and value. We show that consistency applied to the covariant block is provably harmful, shrinking legitimate view-specific content to a fraction $1/(1+4λ)$ of its true value, and we verify this shrinkage law in controlled experiments. Selective cross-view consistency (SCVC) therefore constrains only the invariant block, requires no camera labels, extrinsics, depth, or view synthesis at training or test time, and leaves the deployment interface unchanged. We introduce a carve-and-hold-out evaluation protocol on the LIBERO-Plus camera track that separates a distribution-matched ceiling from genuine interpolation and extrapolation to held-out viewpoints, with a matched pair-trained control isolating the effect of the consistency term from pair exposure. On held-out orbital viewpoints beyond the training envelope, SCVC improves closed-loop success over the matched control by 12.2 points (95% CI [7.4, 17.0]; +15.5, CI [11.7, 19.4], under an independent second seed) -- an effect two further camera axes replicate -- while interpolation within the envelope shows no gain in either seed (-1.2 and -4.3 points) and in-distribution competence is preserved (-0.6, -0.2). We also report a cross-backbone audit showing that published camera-robustness numbers are confounded by wrist-camera pose stability.
Existing adaptive-inference and world-action-model systems use cheap-stage outputs or predicted futures to allocate additional computation. We study a narrower question: under paired exact-reset physical outcomes, can a Medium-derived interface predict when switching to a separately frozen Full predictor improves task-specific decision loss enough to justify sequential overhead? Our contribution is a paired evaluation and audit protocol, not a new generic routing rule: all candidate actions are executed from the same reset state, Medium and Full act on the same candidate set and task, and their paired physical-loss difference defines the routing target. On a fresh PushT bank (V106; 1,600 states, 39 tasks, three checkpoint pairs), a frozen prediction-interface router lowers overhead-inclusive decision cost relative to standalone Medium, standalone Full, and a latency-advantaged task-only router. We then prospectively seal a second 1,600-state PushT confirmation (V107) against a stronger current-state control using the task, a dimension-matched projection of current DINO features, and all five candidate actions, with no DINO encoder latency charged. The prediction interface lowers priced physical decision cost by 0.002549 (state-clustered 95% interval [-0.002867, -0.002238]; one-sided 95% upper bound -0.002286), with negative effects for all three checkpoint pairs. A controlled-PyBullet audit independently supports a composite task-prediction-regime router. The sequential router remains slower than fixed policies, and its advantage is restricted to low compute prices. The evidence supports incremental routing information in the tested prediction interface beyond one deliberately favoured current-DINO control, but not causal sufficiency, compute saving, closed-loop value, or cross-family generality.
Understanding the generalization capabilities of visuomotor policies is essential in the development of capable robotic agents. Generalizable models learn structures that transfer across domains. However, in practice, visuomotor policies test performance by interpolation on known distributions using unstructured domain shifts (e.g. lighting, clutter, diverse objects). We argue that to measure generalization capabilities we must instead test the inductive capacity of policies on progressively harder, out-of-distribution task variants. We call this inductive generalization, drawing directly on how axis-based evaluation has revealed inherent generalization limitations in language models (e.g. sequence length, counting) arXiv:2502.00197 . We provide a reusable and formal evaluation protocol for measuring inductive generalization in any manipulation policy, and establish baselines showing that existing paradigms fail this test; e.g. SoTA Vision-Language-Action models and find that policies that appear to generalize to prior domain shifts (distractors, etc) fail inductive generalization tests. These results expose a class of learning challenges orthogonal to those addressed by data and model scaling in robot learning, yet are imperative to solve in order to realize general purpose robots.
Embodied Vision-Language-Action (VLA) models are typically obtained by fine-tuning powerful pretrained VLMs on robotics data, yet it is unclear how much commonsense and factual knowledge they retain after adaptation. Failures on knowledge-sensitive tasks are ambiguous, conflating missing knowledge with poor generalization of low-level control. We introduce Act2Answer, a lightweight protocol that adapts VLM knowledge benchmarks to VLA evaluation by requiring agents to answer through action. Each question becomes a short tabletop episode where the agent performs a single object-placement action to select among candidate answers, yielding an action-grounded success rate with reduced control confounds. We curate a test suite of such environments across diverse commonsense and world-knowledge categories and introduce layerwise intent probing to localize answer-relevant information across the VLM backbone and action head. In a large-scale study of 7 VLA models and 9 VLM baselines, we systematically rank models across categories, finding that VLAs show solid performance on simple concepts while exhibiting larger gaps on richer semantic categories relative to their source VLMs, that VQA co-training is associated with better knowledge retention, and that answer-relevant signals peak in middle VLA layers but attenuate in upper layers. Act2Answer is available at https://tttonyalpha.github.io/act2answer/.