{"id":"7b8fe4ae-c080-43ce-b5c2-3a19c41b69ba","arxiv_id":"2506.13679","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ROSA trains a VLA model jointly on expert actions and automatically recorded robot states, improving success rates and generalization, particularly with few demonstrations.","lead":"ROSA adds an automatically collected robot-state prediction task to the training of vision-language-action (VLA) models, and the authors report higher task success rates, especially when expert demonstrations are scarce. The method attacks the expensive human data collection bottleneck in generalist robot learning.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim is underdetermined: the random state data is never tested in an action-prediction form, so the reported gains may come from extra random data rather than from the state-estimation supervision itself.","rationale":"The reader's weakest assumption concerns distributional compatibility of the random state data at the 1:4 ratio. That is a real issue, but it presupposes that the benefit is caused by the state-estimation objective; my concern is upstream: the experimental design never isolates the state-estimation label from the additional random data. Both concerns are about what the state data is doing, but they are distinct. The missing control is concrete and feasible, since the same random trajectories contain both current states and next states, so a relabeling experiment can settle it directly. Because the paper currently does not provide this control, the causal claim 'state estimation improves VLA performance' is not fully established; however, the reported empirical results are consistent with the stated mechanism, so the appropriate disposition remains conditional rather than reject or accept. No change to the reader's verdict is needed, but the condition for acceptance should include this control experiment.","tokens_in":14696,"tokens_out":4931,"duration_ms":60111,"concrete_test":"Run the RLBench 100-demo condition with the same random state-collection data and the same 1:4 mixing ratio, but relabel each state sample as an action-prediction sample: use the expert action instruction and set the target to the next recorded end-effector pose from the random trajectory instead of the current pose. Evaluate over the same 25 episodes per task, three repeats, and compare the average success rate to baseline (52.3) and ROSA (63.7). If the relabeled variant matches 63.7 within noise, state-estimation supervision is not the active ingredient; if it stays near 52.3, the state-estimation task is confirmed as the cause.","verdict_should_be":"UNCHANGED","load_bearing_attack":"ROSA's central claim is that supervising current-state estimation is what improves VLA action accuracy. But the training mixture in Sec. 3.1 compares expert action data alone against expert action data plus random-movement data labeled with current states. The random data differs from expert data in two ways at once: trajectory distribution (random vs expert) and target-label semantics (current state vs next action, plus a different instruction). Nothing in Sec. 4 separates these factors. Because the random trajectories are recorded as time series, the same collected observations can be relabeled as action-prediction data by shifting the end-effector pose one step forward and using the expert instruction. If this relabeled variant also gives the large gains in Tab. 1 (e.g., 63.7 at 100 demos), then the improvement is caused by broader coverage of the action space by additional random data, not by the proposed state-estimation objective, and the paper's mechanism story would be unsupported. The linear-probing result in Tab. 7 shows only that ROSA features encode end-effector pose better; it does not identify whether the state-estimation label or the extra random data is responsible. This missing control is the weakest link in the causal claim, more so than the hand-tuned 1:4 ratio, because the 1:4 ratio can be tuned even if the active ingredient is merely extra data.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ROSA, a training paradigm for vision-language-action (VLA) models that adds robot state estimation data to standard expert action data. The state estimation data is collected automatically by letting a robot perform random movements within a constrained scene, recording observations paired with the current end-effector pose and gripper status under a fixed language instruction. This data is mixed with expert demonstrations in a 1:4 ratio and used to fine-tune a Qwen-2.5-7B based VLA model with a unified next-token-prediction loss. The authors report consistent improvements over a baseline on 12 RLBench tasks at five data scales, with the largest gains at 50 and 100 demonstrations per task (7.1 and 11.4 percentage points, respectively), plus real-robot results on seen and unseen tasks, one-shot results, ablations on the state-data ratio and scene relevance, and a linear-probing analysis of 3D understanding.","tokens_in":14990,"tokens_out":2850,"duration_ms":28719,"significance":"If the central claim is established, the paper would be a practical contribution to data-efficient VLA training: it shows that cheap, label-free state supervision can improve action prediction at low data regimes, with controlled comparisons at multiple scales, real-robot generalization experiments, and a credible architectural story motivated by the spatial and temporal gaps in VLM-to-VLA adaptation. The automated data collection pipeline and the explicit comparison with prior methods on RLBench are useful for the community. However, the causal attribution of the gains to the state-estimation objective rather than to the extra random data is not yet demonstrated, and this is the main weakness of the paper.","major_comments":[{"comment":"The reported gains conflate two interventions: the robot state data differs from expert action data in trajectory distribution (random vs. expert), target-label semantics (current state vs. next action), and language instruction. Because the random trajectories are recorded as time series, the same observations can be relabeled as action-prediction data by shifting the end-effector pose one step forward and using the expert instruction. Without this control, the improvements in Tab. 1 could be caused by adding diverse random action data rather than by the current-state estimation objective, which is the paper's stated mechanism. Please add this relabeled-data control, or an equivalent ablation that isolates the label semantics while holding the data distribution fixed, for at least the 50- and 100-demonstration settings.","section":"Sec. 3.1 and Tab. 1"},{"comment":"The linear-probe analysis cannot discriminate between the two factors in the training-mixture confound. ROSA is trained on the state-estimation objective, and the probe is trained on the same objective, so it is unsurprising that ROSA's features decode pose better; the table does not show whether the benefit comes from the auxiliary objective or from the extra random coverage of the action space. Interpret this result as a mechanism illustration rather than causal evidence, or couple it with the relabeled-data control.","section":"Sec. 4.2.4 and Tab. 7"},{"comment":"Real-robot results are reported without error bars or trial-level variance, and the one-shot results in Tab. 2 are reported without variance or evaluation details. Given that the paper's headline real-world claim of a 35% average success-rate improvement rests on these curves, please report the number of evaluation runs, per-seed or per-trial breakdowns, and standard deviations for the real-robot experiments and the one-shot experiments.","section":"Sec. 4.1 and Fig. 5"}],"minor_comments":[{"comment":"There is a typo in the second contribution: 'creat' should be 'create'.","section":"Sec. 1, Contributions"},{"comment":"The table header reads 'Liner-prob evaluation' and should be 'Linear-prob evaluation'.","section":"Sec. 4.2.4, Tab. 7"},{"comment":"The sentence following Eq. (5) says 'ω demotes the parameters' but should read 'denotes'.","section":"Eq. (5)"},{"comment":"'epoches' should be 'epochs' in both places.","section":"Sec. 3.3 and Sec. A.3"},{"comment":"Tab. 6 refers to 'relevant' and 'irrelevant' scenes, but these terms are only defined later in Sec. 4.3; please define them at first use.","section":"Sec. 4.2.4 and Sec. 4.3"},{"comment":"The 'feasible action space' for random state-data collection is described only qualitatively; please specify how the bounds are chosen and how collisions with scene objects are avoided, since this affects the distribution of the collected state data.","section":"Sec. 3.1 and Fig. 2"},{"comment":"The four unseen real-robot tasks are evaluated over 10 trials, but the paper does not state whether the same trials are used across methods or how object placements vary; please clarify the evaluation protocol.","section":"Sec. 4.2.2 and Tab. 3"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"ROSA is a well-executed, clearly written empirical study that shows real gains from adding automatically collected current-state supervision to VLA fine-tuning in low-data regimes. What it doesn't show is that the state label is the active ingredient; the random trajectories are never tested in action-prediction form. That missing control keeps the paper from fully supporting its mechanism claim.\n\nThe actual contribution is a training recipe: mix expert action data with cheap random-movement data labeled with current end-effector pose and gripper state, using a uniform instruction, and get large improvements at 50–100 demos on RLBench (7.1 and 11.4 points) and a big jump on a real WidowX. If that recipe holds up, it's genuinely useful—it attacks the human-data bottleneck without needing extra annotation. The paper deserves credit for the controlled data-scale sweep, the ablations on state-data ratio and scene relevance, and the real-robot generalization tests. The writing is clear and the comparisons are mostly fair; the comparison to LLARVA at 8x data is appropriate.\n\nThe soft spot is exactly what the stress-test says: the random data is new in two ways at once. It's non-expert trajectory distribution and it's labeled with current state. To claim the state-estimation objective matters, the authors need to relabel a random trajectory as action data (take the next pose as the target, keep a normal action instruction) and train on that. If that variant also produces the Tab. 1 gains, the effect is just broader data coverage, not spatial self-awareness. The linear probe in Tab. 7 doesn't settle it—of course features trained to predict end-effector pose encode it better. This is a load-bearing missing control for the mechanism story. The empirical recipe might still be worth publishing, but the paper currently overclaims.\n\nMinor issues: real-robot results have no error bars, one-shot results have no variance, the 1:4 mixing ratio is empirical tuning with no explanation, and no code or data is released. The 'doubles' phrasing in the abstract is selective but not dishonest—it describes a specific real-robot regime.\n\nWho's this for: VLA researchers and anyone working on data-efficient robot learning. It deserves a serious referee—the main experiment is straightforward, and a reviewer can easily ask for the relabeled control. I'd bring it to reading group, but I wouldn't cite it as established until the control is run.","headline":"Useful low-data VLA recipe undermined by a missing control: random data is never tested as action data, so the state label's causal role is unproven.","tokens_in":15543,"tokens_out":2773,"would_cite":false,"duration_ms":29983,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Auxiliary robot state estimation, trained with random-motion data, narrows the vision-language-to-action alignment gap and delivers its largest gains when expert demonstrations are scarce.","keywords":["vision-language-action models","robot state estimation","auxiliary supervision","data efficiency","spatio-temporal alignment","robotic manipulation","RLBench","generalization"],"falsifier":"Collect state data from a feasible space deliberately chosen to avoid the regions of the action space that expert demonstrations occupy (for example, only the far corners of the workspace) and train the same model with the same 1:4 ratio; if the success rate stays at or above the expert-only baseline, the benefit is not due to distributional overlay, and if it collapses, the assumption of distributional compatibility is falsified.","tokens_in":14514,"feed_emoji":"🤖","tokens_out":6097,"duration_ms":65103,"temperature":0.7,"pith_summary":"This paper tries to show that the alignment problem in vision-language-action models can be attacked with an auxiliary task that costs almost nothing. Instead of only fine-tuning a VLM on expert demonstrations of what the robot should do next, ROSA also asks the model to estimate the robot's current 3D end-effector pose and gripper state from the same image. Because the state has the same format as the action, one next-token-prediction objective supervises both, and the state data can be collected automatically by letting the robot move randomly. The paper reports large gains in low-data regimes (7.1 and 11.4 percentage points on RLBench with 50 and 100 demos) and a near-doubling of real-robot success rates, suggesting that cheap self-state supervision is a practical route to data-efficient VLA training.","feed_headline":"Robot states as free training data lift low-data VLA success","feed_subtitle":"Auxiliary pose-estimation data, collected by random robot motion, improves policies most when expert demos are scarce.","key_machinery":"The key mechanism is an auxiliary robot state estimation task built on format-identical data. ROSA records pairs of an image, the uniform language instruction 'What is the current state of the robot?', and a 7-degree-of-freedom state (position, Euler angles, and gripper open/close) by commanding random robot motions within a hand-defined feasible action space, then mixes them with expert action pairs at a 1:4 ratio. Because actions and states share the same 7-dimensional format, the same linear quantization and autoregressive cross-entropy loss handles both, so the model learns to emit a current state and a next action from the same visual input. Ablations show that the 1:4 mixing ratio and a moderate number of scenes matter, while the relevance of the scene content does not.","core_discovery":"The central claim is that a VLA model trained jointly on expert action prediction and robot state estimation outperforms the same model trained on expert data alone, with the largest margin when expert data is scarce. The paper attributes this to closing a spatio-temporal gap: VLM features are semantic and present-oriented, while actions are 3D and future-oriented; state estimation forces the model to produce fine-grained 3D self-awareness, which then supports action prediction. The evidence includes a linear-probe test where ROSA reaches 92% accuracy on 3D understanding versus 61% for the baseline, and real-world generalization trials where average success on unseen tasks rises from 43% to 85%.","pith_inferences":["If the mechanism is mostly distributional coverage rather than self-awareness, then the random state data could be replaced by any cheap off-policy or synthetically sampled action-covering data; a direct test would compare same-format pseudo-actions sampled from a fitted action prior against the recorded random states.","The paper shows scene relevance does not matter for state collection, so the state data could be collected in empty or unrelated environments; this opens a path to cross-embodiment or cross-camera transfer if the fixed camera-robot pose assumption is relaxed.","Because the gains shrink as expert data grows, a natural extension is a data-mixing schedule that anneals the state-data ratio down as expert demonstrations accumulate, which the paper does not explore.","The single fixed viewpoint limitation noted in the appendix suggests a testable extension: collecting state data under varying camera positions could create a self-supervised spatial-consistency objective, potentially improving spatial reasoning further."],"forward_implications":["With 50 or 100 expert demonstrations per task on RLBench, ROSA improves average success rate by 7.1 and 11.4 percentage points over the same baseline, showing that the auxiliary task is most valuable exactly where data is scarce.","On the real robot, ROSA roughly doubles the average success rate in low-data settings and raises average generalization success on unseen tasks from 43% to 85%, suggesting improved robustness to novel objects, containers, and distractors.","In one-shot training, ROSA achieves non-zero success rates on three RLBench tasks where the baseline scores zero on all of them, indicating that state estimation can bootstrap action prediction from a single expert sample.","The benefit persists even at 500 demonstrations per task (a 1.6-point gain), so the auxiliary state task does not merely substitute for data but continues to help when data is abundant, though the marginal gain shrinks.","Adding too much state data degrades performance: at a 1:2 state-to-action ratio the success rate falls from 63.7% to 58.7%, which implies the alignment benefit comes from a careful balance rather than simply more state supervision."],"supporting_citations":[{"why":"Supplies the LLaVA-style vision-language architecture (LLM, vision encoder, projector) that ROSA builds on.","marker":"[1]"},{"why":"Defines the direct fine-tuning paradigm on expert demonstrations that ROSA claims is data-inefficient.","marker":"[10]"},{"why":"Provides the data generation pipeline and keyframe extraction used to build the expert datasets in RLBench.","marker":"[14]"},{"why":"Supplies the CLIP ViT-L/14 vision encoder used in ROSA.","marker":"[25]"},{"why":"The VLA comparison baseline (LLARVA) that ROSA outperforms with eight times less data.","marker":"[29]"},{"why":"Closest prior work using auxiliary spatial annotations to improve data efficiency; ROSA distinguishes itself by needing no extra labels.","marker":"[31]"},{"why":"Supplies the Qwen-2.5-7B large language model backbone used in ROSA.","marker":"[45]"},{"why":"The RLBench simulation environment used for the main success-rate and ablation experiments.","marker":"[46]"}],"fun_headline_variants":["ROSA: state-estimation aux task closes VLA spatio-temporal gap","Robot self-awareness via state estimation lifts VLA data efficiency","State estimation as free supervision improves VLA with scarce demos","Auxiliary robot state estimation boosts VLA action prediction in low-data sets","ROSA: robot state data as extra task for VLA alignment"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on the assumption that random state data collected inside a hand-defined feasible action space, mixed with expert data at a 1:4 ratio, is similar enough to expert actions that the auxiliary task improves action prediction instead of corrupting it.","fun_headline_variants_meta":{"raw":{"variants":["ROSA: state-estimation aux task closes VLA spatio-temporal gap","Robot self-awareness via state estimation lifts VLA data efficiency","State estimation as free supervision improves VLA with scarce demos","Auxiliary robot state estimation boosts VLA action prediction in low-data sets","ROSA: robot state data as extra task for VLA alignment"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000784,"raw_usage":{"total_tokens":3433,"prompt_tokens":887,"completion_tokens":2546,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":2454}},"tokens_in":503,"tokens_out":2546,"duration_ms":19151,"temperature":1.0,"reasoning_tokens":2454,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:27:25.201019+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect state data from a feasible space deliberately chosen to avoid the regions of the action space that expert demonstrations occupy (for example, only the far corners of the workspace) and train the same model with the same 1:4 ratio; if the success rate stays at or above the expert-only baseline, the benefit is not due to distributional overlay, and if it collapses, the assumption of distributional compatibility is falsified.","supporting_citations":[{"cited_title":"Llarva: Vision-action instruction tuning enhances robot learning","cited_arxiv_id":null,"evidence_quote":"The VLA comparison baseline (LLARVA) that ROSA outperforms with eight times less data."},{"cited_title":"push the maroon button, then push the green button","cited_arxiv_id":null,"evidence_quote":"The RLBench simulation environment used for the main success-rate and ablation experiments."}],"review_version":1}