{"id":"b9de3beb-b2af-4cd0-a9be-302e68058666","arxiv_id":"2505.12072","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"L2D2 lets humans teach robot tasks by drawing on synthetic images of varied scenes and adding a few physical corrections, achieving teleop-level policy performance with less user effort.","lead":"This paper presents L2D2, a system for teaching robots by drawing on images instead of physically moving the robot. In tests, it matches teleoperation performance while needing far fewer physical demonstrations and less user time.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 5.3 reports t(18)=11.37 with p=0.441, an impossible combination; the long-horizon 'as performant as Teleop' result is therefore unsupported as written.","rationale":"The reader's verdict is CONDITIONAL and its weakest-assumption is the physical validity of the synthetic drawing images. I agree that this is an open gap, but I find a more immediately checkable problem in the reported statistics. The strongest_claim is presented as a general equivalence result, and the abstract specifically claims generalization to longer-horizon tasks; Section 5.3 is the only experiment supporting that generalization, and its only inferential evidence is internally inconsistent. A single impossible t/p pair is easy to fix typographically, but until it is corrected the long-horizon arm of the central claim has no valid support. This does not overturn the rest of the paper: the short-horizon comparisons in Section 5.1 and the user study time/effort results are reported with plausible ANOVA and post-hoc statistics, and the method is coherent. The concern is therefore not a rejection; it is a request to correct or provide the raw data for one decisive result. The concrete test above settles it. If the corrected p remains non-significant, I would accept the conditional verdict as is; if not, the long-horizon generalization claim should be narrowed. Hence verdict_should_be is UNCHANGED relative to the reader's CONDITIONAL, with the correction as a condition.","tokens_in":22520,"tokens_out":11473,"duration_ms":118569,"concrete_test":"Request the raw per-rollout success rates (or corrected t and p) for Section 5.3's Long Horizon experiment and recompute the independent-samples t-test with df=18. If the corrected p exceeds 0.05 with a consistent t, the equality claim stands and only the typo needs correction. If the corrected p is below 0.05, or if the corrected test reverses the conclusion, Section 5.3's 'similar performance' statement must be revised and the long-horizon generalization claim in the abstract adjusted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's broadest claim—that L2D2 can match full teleoperation while using only 10 physical demonstrations—is explicitly extended to long-horizon tasks in Section 5.3. The only statistical support offered for that extension is 't(18)=11.37, p=0.441.' These two numbers cannot both be correct: with 18 degrees of freedom, t=11.37 corresponds to p<0.0001, whereas p=0.441 corresponds to |t|≈0.79. The paper reports no means, SEMs, or effect sizes for the Long Horizon rollouts, so the reader cannot determine whether L2D2 was actually equivalent, better, or worse than Teleop in this condition. Because the abstract and conclusion use the long-horizon result to claim generalization beyond short single-object tasks, the central claim currently rests on a statistically invalid sentence. This is most plausibly a transcription error, not evidence of misconduct, but as printed the evidence cannot be checked. A corrected statistic, or the raw rollout data, is needed before the long-horizon equality claim can be accepted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes L2D2, a sketching interface and imitation learning pipeline in which human users draw end-effector trajectories, orientations, and gripper commands on images of the robot workspace. A vision-language detector segments task-relevant objects, and the system generates synthetic images with varied object placements so that users can draw diverse demonstrations without physically resetting the environment. A task-agnostic neural mapping converts 2D sketch points to 3D positions, object dynamics are approximately simulated, and a small set of physical demonstrations is used to fine-tune both the mapping and the behavior-cloned policy. The manuscript reports short-horizon and long-horizon tabletop experiments with expert users, a 12-participant user study comparing L2D2 with teleoperation and two sketch-based baselines, and a PCA-based camera placement analysis.","tokens_in":22752,"tokens_out":6465,"duration_ms":60378,"significance":"If the results hold, L2D2 addresses a real bottleneck in imitation learning: the cost of collecting diverse physical demonstrations. The reported short-horizon experiments and user study support the claim that drawings plus a few physical corrections can match teleoperation on single-object tasks while reducing user time and effort, and the subjective results are consistent with that conclusion. The PCA-based camera placement derivation is internally coherent and has no fitted parameters, and the paper makes its code and project materials available, which are strengths. The ablations (L2D2-D and Teleop-min) are well chosen and help separate the contributions of diverse drawings and physical grounding. However, the long-horizon generalization claim in the abstract and conclusion is currently supported only by an internally inconsistent test statistic, so the overall evidence for the broadest claim is incomplete.","major_comments":[{"comment":"The reported test statistic is internally inconsistent: for an independent-samples t-test with 18 degrees of freedom, t = 11.37 corresponds to p < 0.001 (two-tailed), while p = 0.441 would require |t| ≈ 0.79. Because no means, SEMs, or effect sizes are reported for the Long Horizon rollouts, the current text does not establish that L2D2 and Teleop performed similarly. Since the abstract and conclusion explicitly extend the claim to longer-horizon tasks, this is a load-bearing issue; please correct the statistic or provide the raw data and effect sizes.","section":"§5.3, Results"},{"comment":"The inequality Inonlinear_loss(C) ≤ Iloss(C) is presented as a consequence of using a nonlinear mapping, and the abstract says that an upper bound is established, but no proof or precise formal statement is given. The cited prior work concerns autoencoders versus PCA in general, not the specific perspective-projection reconstruction problem here; moreover the inequality cannot hold for arbitrary learned f_phi without assumptions about the function class and training optimality. Please either provide a proof for the stated setup or reclassify this as an empirical observation with supporting experiments.","section":"§4.2, Eq. (9)"},{"comment":"The camera placement analysis in Section 4.2 is a stated contribution, but no experiment varies the camera position to test whether the PCA-optimal placement improves reconstruction accuracy or policy success relative to alternative placements. The paper only says that the camera was placed 'as best as we can'. A comparison (e.g., PCA-optimal versus a fixed non-optimal viewpoint, holding all else equal) would be needed to validate Eq. (7) empirically.","section":"§5, Experimental Setup"},{"comment":"The synthetic-image pipeline randomly repositions object masks and inpaints the vacated regions, but the manuscript does not validate that the resulting scenes are physically plausible or that user drawings on synthetic images are as accurate as drawings on real images. The assertion that 'moving objects to random image locations introduces sufficient variety' is not supported by quantitative evidence. Since the diversity claim is central to the data-collection contribution, please validate this step (e.g., by comparing drawing accuracy or policy success when synthetic images are replaced by real images with the same object placements, or by filtering implausible placements).","section":"§4.1, Diverse Images"},{"comment":"The RT-Traj baseline is simplified by feeding sketched trajectories directly instead of sketch images as in the original paper, and both RT-Traj and S2S use 160 task-specific demonstrations to train the mapping. This may be a reasonable data-budget choice, but as described it weakens the claim that L2D2 outperforms the state-of-the-art sketch-based methods, because the comparison may be against degraded versions of those methods. Please justify that the simplified baselines are representative, or report their original-form performance where feasible.","section":"§5.1, Baselines"}],"minor_comments":[{"comment":"The text says 'Similar to Section 5.2' when describing the evaluation; this should refer to Section 5.1.","section":"§5.2, Dependent Variables"},{"comment":"The intuitiveness ANOVA is reported as 'F (2.01, 24.1), p < 0.05' without the F-statistic value; please insert it.","section":"§5.2, Results"},{"comment":"Post-hoc comparisons are reported without a stated multiple-comparison correction; please state the procedure used (e.g., Tukey, Bonferroni, or the specific method embedded in the software).","section":"§5.1 and §5.3"},{"comment":"The axis label 'rotxrotyrotz' appears to be a concatenation of rotx, roty, and rotz; format it clearly for readability.","section":"Figure 2"},{"comment":"Reference [24] (Jolliffe and Cadima) has incomplete bibliographic information; please add the volume, article number, or page range.","section":"References"},{"comment":"The notation P_C is used both for the set of robot states in the camera frame and for the data matrix in Eq. (6); please distinguish the two to avoid confusion.","section":"§4.2, Equations (5)-(7)"}],"recommendation":"major_revision","confidential_remarks":"The main reason for major_revision is the impossible t-statistic in Section 5.3 combined with the paper's reliance on that result for the long-horizon generalization claim. If the corrected statistic turns out to be non-significant with a small effect, the authors should soften the abstract and conclusion accordingly. I also see a risk that the sketch-baseline comparisons may be against simplified implementations, which is worth checking carefully. Overall, the short-horizon evidence is solid and the work fits the journal's scope; the requested changes are local enough that I expect the authors can address them in a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth your time. L2D2 combines sketch-based teaching with VLM-generated synthetic image variation and a small set of physical corrections. The short-horizon experiments genuinely support the headline claim: L2D2 matches teleoperation performance with roughly one-sixth the physical demonstrations, and the user study shows users spend less time and prefer the interface. The camera-placement PCA analysis is a nice theoretical addition, even if it is not directly ablated.\n\nThe main problem is in Section 5.3. The long-horizon comparison reports t(18)=11.37, p=0.441, which cannot both be true. With 18 degrees of freedom, t=11.37 gives p well below 0.0001; p=0.441 corresponds to |t| around 0.79. The paper gives no means or SEMs for that condition, so the equality claim rests on an invalid statistic. The abstract and conclusion use that result to argue L2D2 generalizes to long-horizon tasks, so this needs a correction or raw rollout data before that claim is usable. I believe it is a transcription error rather than anything sinister, but as printed it is evidence that cannot be checked.\n\nOther soft spots are more minor. The RT-Traj and S2S baselines are simplified in ways that may be unfair: they get direct trajectories instead of image-conditioned input, do not receive rotation annotations, and require 160 task-specific demos to train their mappings. The synthetic image generation randomly repositions object masks without checking physical plausibility; the paper relies on physical corrections to clean that up, which mostly works in the short tasks but is worth a sensitivity check. The claim on nonlinear mapping upper bound (Eq. 9) is asserted without proof, and the camera-placement optimization is not experimentally validated—though both are secondary.\n\nNone of this undermines the short-horizon result, which is consistent across two tasks and both expert and novice data. The paper is clearly written, the limitations are acknowledged honestly, and the code is available. I would send it to peer review, with the long-horizon statistic flagged as a required fix. Researchers in imitation learning and HRI should engage with this work; it is a genuine step toward lowering the cost of diverse demonstrations.","headline":"Short-horizon results genuinely support L2D2's main claim; the long-horizon result is currently unsupported because of an impossible t-test statistic.","tokens_in":23316,"tokens_out":3144,"would_cite":true,"duration_ms":29828,"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":"L2D2 argues that teaching robot arms by drawing on synthetically varied images, then grounding those drawings with a few physical corrections, yields policies as effective as full teleoperation while requiring far less human time and…","keywords":["learning from demonstration","sketch-based teaching","2D to 3D mapping","synthetic data augmentation","behavior cloning","human-robot interaction","camera placement","vision-language segmentation"],"falsifier":"Run a controlled comparison where the same set of object poses is presented either by physically resetting the workspace or by L2D2's random mask repositioning; if policies trained on drawings from synthetic images are systematically worse than those trained on drawings from real images at the same poses, or if users' drawings on synthetic images frequently intersect objects or leave the table, the claim that synthetic drawings are sufficient would be refuted.","tokens_in":22296,"feed_emoji":"🖍️","tokens_out":7116,"duration_ms":67174,"temperature":0.7,"pith_summary":"L2D2 is a sketching interface and imitation learning pipeline that lets a human teach a robot arm by drawing trajectories on images of the workspace instead of physically guiding the robot. The paper's central claim is that combining many such drawings, collected quickly on synthetically varied images, with a small set of physical demonstrations produces robot policies as effective as those trained on much larger sets of full teleoperated demonstrations. In the reported experiments, 50 drawings plus 10 physical corrections matched 60 teleoperated demonstrations on Lift and Push tasks, and similar results appeared in a novice user study and a two-subtask long-horizon task. The reason this matters is that drawings remove the most costly parts of conventional data collection: resetting the environment between examples and physically operating the robot. If the claim holds, scaling up demonstration data for manipulation no longer requires proportional human physical effort.","feed_headline":"Drawings plus 10 demos match 60 teleop demos","feed_subtitle":"L2D2 turns sketching on synthetic images into robot policies that match teleoperation, with far less human effort.","key_machinery":"The central object is the L2D2 pipeline, which has four load-bearing parts. The sketching interface lets a person draw a 2D path for the end-effector, attach orientation and gripper open/close commands at selected points, and label task-relevant objects through text prompts. A camera-placement rule, derived from a PCA variance model, positions the camera so that the depth axis carries as little of the robot's motion variance as possible, minimizing the information lost when 3D states are projected to 2D pixels. A neural network $f_\\phi$ maps drawn pixel points back to 3D robot positions, trained on task-agnostic pairs generated by projecting uniformly sampled workspace points. Finally, a vision-language object detector segments the labeled objects and repositions them in the image, so users can draw many demonstrations of the same task in varied scenes; a small physical dataset is then used to fine-tune $f_\\phi$ and the behavior cloning policy sequentially. The information-loss identity $I_{\\mathrm{loss}}(C)=\\mathrm{Var}(z_C)/(\\mathrm{Var}(x_C)+\\mathrm{Var}(y_C)+\\mathrm{Var}(z_C))$ is what connects camera placement to reconstruction quality.","core_discovery":"The paper claims that a robot can learn manipulation tasks from static, 2D drawings on synthetic images as effectively as from physical demonstrations, provided the drawings are grounded by a small real-world dataset. The grounding works in two steps: physical demonstrations are used to fine-tune the 2D-to-3D mapping, so that the drawing dataset is reconstructed more accurately, and then the policy is trained sequentially on the reconstructed drawings and the physical data. In the experiments, policies trained with 50 drawings and 10 physical corrections were statistically indistinguishable from policies trained with 60 teleoperated demonstrations in short-horizon tasks, and comparable performance held in a novice user study and a two-subtask long-horizon setting. The intended contribution is a demonstration that low-information, static sketches can be converted into dynamic, high-dimensional robot behavior with only a few real-world interventions.","pith_inferences":["The information-loss analysis implies a general design rule for any sketch-based teaching system: position the camera so that task-relevant motion is spread across the image plane rather than along the viewing axis, a rule that could transfer to other interfaces without L2D2's other components.","Random mask repositioning likely works when the task is roughly planar and the physical correction set is diverse enough to correct errors; tasks with stacking, containment, or coupling between objects would probably need physics-aware augmentation or filtering.","The gripper-based heuristic for simulating object motion is a simplification; replacing it with learned video prediction or physics simulation is a natural testable extension that could reduce the number of physical corrections required.","A direct stress test would be to increase the distance between training and test object distributions to see whether the diversity from synthetic drawings generalizes beyond the range of object positions covered by the physical corrections."],"forward_implications":["Teaching effort and time scale more favorably because environment resets are replaced by synthetic image variation, so the human cost per additional demonstration is mostly the time to draw.","A small amount of physical data can anchor a much larger drawing dataset, since the physical corrections are used to recalibrate the 2D-to-3D mapping and then to fine-tune the policy trained on reconstructed drawings.","Drawing-based teaching is not limited to short single-object tasks: the long-horizon experiment shows comparable success to teleoperation when the task consists of two sequential subtasks.","User preference and perceived effort shift in favor of drawing interfaces: in the user study, most participants preferred drawing over teleoperation and preferred L2D2 over the other drawing baselines."],"supporting_citations":[{"why":"Serves as the RT-Traj sketching baseline that L2D2 is compared against.","marker":"[17]"},{"why":"Serves as the S2S two-view sketching baseline and the performance reference for gripper tasks.","marker":"[51]"},{"why":"Provides the vocabulary-based object detector used to segment and reposition objects for synthetic image generation.","marker":"[58]"},{"why":"Underpins the PCA model of information loss and the optimal camera placement objective.","marker":"[24]"},{"why":"Supplies the pinhole camera projection model used to define the 2D-to-3D mapping and its calibration data.","marker":"[18]"},{"why":"Supplies the behavior cloning objective used to train all compared policies.","marker":"[38]"},{"why":"Defines the DAgger-style data merging that L2D2 deliberately avoids with sequential training.","marker":"[41]"}],"fun_headline_variants":["Sketch to teach: 50 drawings + 10 demos match 60 teleops","Robot policies from 2D drawings, grounded by 10 real demos","Draw tasks, don't teleoperate: 50+10 equals 60","L2D2: 50 sketches and 10 physical demos beat teleoperation","Teaching robots by drawing: 50 sketches + 10 demos = 60 teleops"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that randomly repositioning object masks in the image creates scenes where users' drawings are still valid demonstrations of the task; if the synthetic scenes are physically implausible or if users draw inaccurately on them, the large drawing dataset will contain misleading examples that the few physical corrections may not be able to fix.","fun_headline_variants_meta":{"raw":{"variants":["Sketch to teach: 50 drawings + 10 demos match 60 teleops","Robot policies from 2D drawings, grounded by 10 real demos","Draw tasks, don't teleoperate: 50+10 equals 60","L2D2: 50 sketches and 10 physical demos beat teleoperation","Teaching robots by drawing: 50 sketches + 10 demos = 60 teleops"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000687,"raw_usage":{"total_tokens":3167,"prompt_tokens":1051,"completion_tokens":2116,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":667,"completion_tokens_details":{"reasoning_tokens":2006}},"tokens_in":667,"tokens_out":2116,"duration_ms":17702,"temperature":1.0,"reasoning_tokens":2006,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:42:01.066193+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a controlled comparison where the same set of object poses is presented either by physically resetting the workspace or by L2D2's random mask repositioning; if policies trained on drawings from synthetic images are systematically worse than those trained on drawings from real images at the same poses, or if users' drawings on synthetic images frequently intersect objects or leave the table, the claim that synthetic drawings are sufficient would be refuted.","supporting_citations":[{"cited_title":"In: Interna- tional Conference on Learning Representations (2024)","cited_arxiv_id":null,"evidence_quote":"Serves as the RT-Traj sketching baseline that L2D2 is compared against."},{"cited_title":"In: European Conference on Computer Vi- sion, pp","cited_arxiv_id":null,"evidence_quote":"Provides the vocabulary-based object detector used to segment and reposition objects for synthetic image generation."},{"cited_title":"Philosophical trans- actions of the royal society A: Mathematical, Physical and Engineering Sciences","cited_arxiv_id":null,"evidence_quote":"Underpins the PCA model of information loss and the optimal camera placement objective."},{"cited_title":"Cambridge University Press (2003)","cited_arxiv_id":null,"evidence_quote":"Supplies the pinhole camera projection model used to define the 2D-to-3D mapping and its calibration data."},{"cited_title":"Advances in Neural Information Processing Systems 1 (1988)","cited_arxiv_id":null,"evidence_quote":"Supplies the behavior cloning objective used to train all compared policies."},{"cited_title":"In: International Conference on Artificial Intelligence and Statistics (2011)","cited_arxiv_id":null,"evidence_quote":"Defines the DAgger-style data merging that L2D2 deliberately avoids with sequential training."}],"review_version":1}