{"id":"4848fe1b-742e-4c18-96f2-5d5d4bf151bc","arxiv_id":"2412.13548","paper_version":6,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"TelePreview adds a physically aligned augmented-reality preview and a preview/execute foot-pedal switch to low-cost glove-and-IMU teleoperation, and reports higher success rates and shorter execution times in a five-task evaluation.","lead":"TelePreview is a teleoperation system that shows the operator a virtual preview of the robot's next move before the real robot executes it, with a foot pedal to switch between preview and execution. The paper reports that this preview feature improved success rates and speed for new users across five real manipulation tasks, with input hardware costing less than $1,000.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline 'outperforms other teleoperation systems' rests on cross-paper success rates from different setups, not matched baselines; internal tie and negative per-task result further weaken the claim.","rationale":"TelePreview is a plausible system: the AR preview with foot-pedal switching is a sensible interface, the hardware cost is roughly as claimed, and the internal pipeline (SMPL-X abstraction, joint mapping, collision-avoidance networks with FK-generated labels) is technically coherent. The strongest evidence is the within-system comparison of preview on/off in Table II; on average, preview improves success (0.64 to 0.92) and reduces execution time (31.6 to 23.7 s), and the system's stated limitation of alpha-blending occlusion is honestly disclosed. However, the headline claim of outperforming other teleoperation systems is load-bearing and is not supported by the presented data. Table I compares TelePreview's numbers on its own hardware and tasks against independently reported numbers from other papers; those numbers come from different robots, hands, objects, operators, and success criteria. The paper even states it encountered local deployment failures and therefore quotes published results (Section VI-D, Appendix C), which cannot establish comparative superiority. The claim is also internally overstated: AnyTeleop ties TelePreview on Pick & Place (1.0 vs 1.0), and Table II shows Pour success decreasing with preview and Hang time increasing, contradicting the text's per-task claims. These inconsistencies suggest the empirical section was not carefully checked, but they do not negate the system's potential value. The correct remedy is a matched-protocol benchmark or a softened claim. Thus the CONDITIONAL verdict stands; the paper should not be accepted as demonstrating superiority without such a comparison.","tokens_in":13383,"tokens_out":6739,"duration_ms":55857,"concrete_test":"Run AnyTeleop, OpenTeach, and Telekinesis (or their public checkpoints) on the same UFactory xArm + LEAP Hand hardware, the same five tasks, objects, trial count (10), and success criteria used for TelePreview, with the same operators. If TelePreview's success rates are not strictly higher on the matched runs, the claim 'outperforms other teleoperation systems' should be revised to 'is competitive with' or 'is effective in our setup.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim that TelePreview 'outperforms other teleoperation systems across five tasks' depends entirely on success rates quoted from other papers (Section VI-A: 'We reference the baseline success rates from their papers'). These numbers were measured on different robot hands, objects, task definitions, operators, and success criteria, so they are not comparable to TelePreview's numbers in Table I. The paper itself acknowledges local deployment issues (Appendix C) but does not provide matched-protocol baseline runs. Moreover, the claim is internally overstated: Table I shows AnyTeleop also achieves 1.0 on Pick & Place, so TelePreview does not strictly exceed 'all baselines.' In Table II, the Pour task success rate drops from 0.9 (w/o preview) to 0.8 (w/ preview), contradicting the statement that 'success rates increase across all tasks,' and Hang execution time increases by +1.53 s. While the average across tasks favors preview, the per-task claims are not consistently supported. Without a controlled comparison (same robot, tasks, objects, operators, and criteria), the headline 'outperforms' is not established; the evidence only shows TelePreview works well in its own setup.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces TelePreview, a teleoperation system that overlays a spatially aligned virtual robot on the real scene so that an operator can preview and refine commands before physical execution. The system combines IMU-based wrist tracking, mocap glove hand tracking, an SMPL-X abstraction layer, and a learned collision-avoidance retargeting module (CPN/CCN), with a reported hardware cost of under $1,000. The experiments address three questions: (Q1) effectiveness versus prior teleoperation systems, (Q2) benefit of the preview for new users, and (Q3) adaptability to different end-effectors. The evaluation includes success rates and execution times on five real-world manipulation tasks, a user study with 5 new users comparing with and without preview, and deployment on a dexterous hand, a parallel-jaw gripper, and a vacuum gripper. The paper claims that TelePreview outperforms other teleoperation systems across the five tasks and that preview assistance boosts new-user success and speed.","tokens_in":13579,"tokens_out":3332,"duration_ms":31608,"significance":"If the claims are established, TelePreview would be a useful, low-cost teleoperation interface that improves data-collection quality by letting operators separate exploration from execution. The paper has several genuine strengths: it releases code and deployment documentation, gives a detailed bill of materials, and runs a same-protocol internal comparison (with vs. without preview) that mostly shows consistent gains in success rate and execution time for new users. The multi-view preview design is a practical contribution, and the adaptation to three end-effectors with minimal parameter changes is well motivated. However, the headline claim of outperforming prior systems is not supported by the presented evidence, because the baseline success rates are quoted from other papers with different hardware, tasks, objects, operators, and success criteria. The internal user study is the strongest evidence, but its per-task claims are not uniformly consistent with the data.","major_comments":[{"comment":"The claim that TelePreview achieves \"higher success rates than all baselines\" rests on success rates quoted from other papers (Open Teach [1], AnyTeleop [3], Telekinesis [32]), as stated in Section VI-A: \"We reference the baseline success rates from their papers.\" These numbers were measured on different robot hands, objects, task definitions, operators, and success criteria, so they are not comparable to the TelePreview numbers in Table I. Without matched-protocol baseline runs on the same hardware and tasks, Table I only shows that TelePreview works well in its own setup, not that it outperforms prior systems. Moreover, the table shows AnyTeleop also achieving 1.0 on Pick & Place, so TelePreview does not strictly exceed all baselines even under the authors' own tabulation.","section":"Section VI-A, Table I"},{"comment":"The statement that \"success rates increase across all tasks\" is contradicted by the Pour task, where the success rate drops from 0.9 (w/o preview) to 0.8 (w/ preview), a difference of -0.1. In addition, the Hang task shows a small increase in execution time (+1.53 s) when the preview is enabled. While the average across tasks favors the preview, the per-task claims are not consistently supported. With only 5 new users and no statistical significance testing, the strength of the Q2 conclusion should be tempered to reflect the observed variability.","section":"Section VI-E, Table II"},{"comment":"The collision-avoidance component (CPN and CCN) is described as a key part of the teleoperation pipeline, but no experiment quantifies its contribution. In particular, there is no ablation showing collision rates during teleoperation with and without the network, nor a comparison of task success with and without the correction step. The appendix reports that baseline methods suffered self-collisions in the authors' reproduction attempts, but this is anecdotal and not a controlled measurement. Without such an evaluation, the \"non-collision retargeting\" property and its role in the reported task success rates remain unverified.","section":"Section IV-C, Appendix A"}],"minor_comments":[{"comment":"The sentence \"But Many existing teleoperation frameworks\" should be \"However, many existing teleoperation frameworks\".","section":"Section II-A"},{"comment":"The motion planning library \"mplib\" is used without a citation or reference; please add the appropriate bibliographic entry.","section":"Section V-B.3"},{"comment":"The header \"Average Success\" is ambiguous; consider renaming it to \"Success Rate (mean)\" to clarify that it is the mean over 10 trials.","section":"Table II"},{"comment":"There are spacing artifacts in the table (e.g., \"13 .6\" and \"14 .2\"); also, the LeapHand row duplicates the Pick & Place numbers from Table II, and the user population (new users vs. the general group) should be stated explicitly in the caption.","section":"Table III"},{"comment":"The captions should identify the \"baseline vision-based method [3]\" specifically as AnyTeleop, and state how many participants provided the practice-time data.","section":"Appendix D, Figures 9-10"},{"comment":"The world frame is defined as the midpoint between the feet, but Eq. (1) uses pw(0) as the initial wrist position; please clarify whether pw(0) and pe(0) are recorded at the same time and how the initial end-effector pose is obtained.","section":"Section IV-A, Eq. (1)"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid engineering contribution, and the internal with/without preview comparison is the most credible evidence. The main weakness is the cross-paper baseline comparison in Table I, which should be either replaced with matched-protocol runs or substantially reframed. I would ask the authors to revise the title and abstract claims to match what the evidence actually supports, and to add the missing ablation for the collision-avoidance module."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead TelePreview. The useful thing is the integrated system: spatially aligned AR preview of the robot overlaid on the physical robot, a foot-pedal switch between preview and execution, and a sub-$1000 glove+IMU rig with SMPL-X retargeting and learned collision correction. That combination is genuinely new; prior AR teleop systems mostly didn't align the virtual arm to the physical one, or didn't have the preview/execute state machine. The internal comparison (preview on vs off, Table II) mostly supports the core claim: new users succeed more often and finish faster with the preview on five tasks. They release code and a deployment doc, which is real.\n\nThe soft spots are in the headline claim, not the system. 'Outperforms other teleoperation systems' rests on success rates quoted from Open Teach, AnyTeleop, and Telekinesis papers, measured on different hands, objects, tasks, and operators. Table I even shows AnyTeleop matching 1.0 on Pick & Place, so the 'exceeds all baselines' phrasing is false on its face. The paper acknowledges local deployment issues (Appendix C) but doesn't run matched-protocol baselines. The new-user study has 5 participants, no significance testing, and the text says 'success rates increase across all tasks' when Table II shows Pour dropping 0.9 to 0.8 and Hang execution time increasing by 1.5s. Those are internal inconsistencies, not just missing stats.\n\nNone of this kills the paper. The preview benefit is still plausible and mostly positive on average. But the central empirical claim needs a controlled benchmark with more participants and re-run baselines on the same setup, or a carefully hedged claim about cross-paper comparison. As is, the paper is a solid systems description with an overblown experimental section.\n\nWorth peer review? Yes. A serious referee should push on the baseline comparison and the per-task inconsistencies, but the system and the internal preview study are substantive enough to warrant revision, not desk rejection. I'd bring it to reading group mainly to discuss how far cross-paper numbers can support 'outperforms.'","headline":"Solid teleoperation system with a genuinely useful preview concept, but the headline outperformance claim is propped up by non-comparable cross-paper numbers and a couple of internal inconsistencies.","tokens_in":14195,"tokens_out":2027,"would_cite":true,"duration_ms":18980,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"TelePreview claims that letting operators visualize and refine a robot's next movement in an aligned virtual preview before physical execution improves teleoperation success rates and safety while keeping hardware under $1,000.","keywords":["teleoperation","virtual preview","augmented reality","dexterous manipulation","imitation learning","SMPL-X","motion retargeting","low-cost robotics"],"falsifier":"Run the same five manipulation tasks on the same robot, with the same objects, same environment, and same five new users, using TelePreview, Open Teach, AnyTeleop, and Telekinesis under an identical protocol; if TelePreview does not achieve higher success rates or shorter execution times than the baselines, the paper's headline comparison fails.","tokens_in":13120,"feed_emoji":"🤖","tokens_out":6241,"duration_ms":50994,"temperature":0.7,"pith_summary":"This paper presents TelePreview, a teleoperation system that overlays a spatially aligned virtual robot on the real scene so an operator can see the outcome of the next movement before the physical robot executes it. The authors aim to show that this preview-then-execute design makes fine-grained manipulation tasks easier for new users, safer for hardware, and cheaper than existing dexterous teleoperation systems, with a hardware cost under $1,000. They report higher success rates than three published teleoperation systems across five real-world tasks, and a user study in which enabling the preview raised new-user success rates and cut execution times. If the claims hold, the system offers a low-cost route to cleaner demonstration data for imitation learning, because exploratory moves are filtered out before they reach the robot.","feed_headline":"Preview-first teleoperation beats baselines on five tasks","feed_subtitle":"A sub-$1,000 system lets operators refine commands on a virtual arm before the real robot moves.","key_machinery":"The central object is the preview overlay: a virtual robot rendered from a 3D model and registered to the physical robot through AprilTag-based hand-eye calibration, then alpha-blended into live camera images from multiple viewpoints. A foot pedal toggles between Preview Mode, where the virtual arm moves with the user's live hand while the physical robot is frozen, and Align Mode, where the final preview pose becomes the target for a motion-planning library and the physical robot executes it autonomously. Supporting this is a teleoperation pipeline that maps IMU-tracked wrist pose and mocap-glove hand joints onto the robot through the SMPL-X body model, a linear joint-to-joint mapping with per-joint scale, bias, and direction, and a learned self-collision prediction/correction network that keeps high-degree-of-freedom hand configurations valid.","core_discovery":"TelePreview's central claim is that inserting a virtual preview step between human command and robot execution—rendered as a robot arm aligned with the physical arm via AprilTag calibration, with a foot pedal toggling between preview-only and align modes—lets operators issue refined, collision-free commands. The paper argues this separation of exploration from execution reduces self-collisions, shortens task times, and improves success rates, and that the underlying SMPL-X based retargeting pipeline makes the system portable across input devices and end-effectors with only small parameter changes. The evidence is success-rate comparisons on five manipulation tasks (Pick & Place, Pour, Hang, Box Rotation, Cup Stacking) against Open Teach, AnyTeleop, and Telekinesis, plus a within-system user study comparing preview enabled versus disabled. The paper frames the preview as a way to capture only intentional, task-oriented motions in demonstration data, avoiding the exploratory movements that degrade imitation-learning policies.","pith_inferences":["The paper does not test bimanual or mobile manipulators, so extending the preview design to those settings is an untested consequence of the architecture.","An apples-to-apples replication with identical operators, hardware, and objects may shrink the reported margin over baselines, because the current comparison uses numbers quoted from different papers.","Replacing alpha blending with depth-aware rendering could automatically resolve the occlusion ambiguities the paper identifies, potentially improving new-user performance further.","The claim that previewing yields cleaner demonstrations rests on the assumption that exploratory motions during preview are not recorded; the paper reports execution-time improvements but does not directly measure downstream imitation-learning accuracy."],"forward_implications":["If the preview works as claimed, new users can produce task-ready demonstrations with less practice, lowering the barrier to collecting imitation-learning data.","Separating exploration from execution yields cleaner demonstration trajectories, which should improve the generalization of learned manipulation policies.","Sub-$1,000 hardware with modular input and output support makes the system feasible for research labs with limited budgets.","The preview-to-execute workflow protects robot hardware from self-collisions and unsafe commands during training data collection.","Multi-view rendering and AprilTag alignment could transfer to other camera-based augmented reality interfaces for robot teleoperation."],"supporting_citations":[{"why":"Provides the Open Teach baseline success-rate numbers used in Table I for cross-system comparison.","marker":"[1]"},{"why":"Provides the AnyTeleop baseline success-rate numbers and the vision-based retargeting approach whose self-collisions motivate the collision-avoidance module.","marker":"[3]"},{"why":"Provides the Telekinesis baseline success-rate numbers and a YouTube-learned hand imitator that the non-collision retargeting improves upon.","marker":"[32]"},{"why":"Supplies the AprilTag fiducial system used to spatially align the virtual preview with the physical robot.","marker":"[35]"},{"why":"Supplies the SMPL-X body model that standardizes wrist and hand pose representation across input devices.","marker":"[36]"},{"why":"Supplies the LEAP hand, the 16-DoF anthropomorphic end-effector used in the main experiments.","marker":"[37]"}],"fun_headline_variants":["Virtual preview arm makes teleoperation safer and faster","Sub-$1,000 teleoperation with virtual arm preview wins five tasks","See the robot's next move before it moves: TelePreview","TelePreview: Preview commands on a virtual arm, then execute","Virtual arm preview boosts teleoperation success on five tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that TelePreview outperforms other teleoperation systems depends on assuming that the success rates quoted from the Open Teach, AnyTeleop, and Telekinesis papers are directly comparable to TelePreview's numbers, even though each was measured on different hardware, objects, and operators.","fun_headline_variants_meta":{"raw":{"variants":["Virtual preview arm makes teleoperation safer and faster","Sub-$1,000 teleoperation with virtual arm preview wins five tasks","See the robot's next move before it moves: TelePreview","TelePreview: Preview commands on a virtual arm, then execute","Virtual arm preview boosts teleoperation success on five tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000704,"raw_usage":{"total_tokens":3178,"prompt_tokens":950,"completion_tokens":2228,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":2145}},"tokens_in":566,"tokens_out":2228,"duration_ms":15006,"temperature":1.0,"reasoning_tokens":2145,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:01:12.989991+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same five manipulation tasks on the same robot, with the same objects, same environment, and same five new users, using TelePreview, Open Teach, AnyTeleop, and Telekinesis under an identical protocol; if TelePreview does not achieve higher success rates or shorter execution times than the baselines, the paper's headline comparison fails.","supporting_citations":[{"cited_title":"Expressive body capture: 3d hands, face, and body from a single image,","cited_arxiv_id":null,"evidence_quote":"Supplies the SMPL-X body model that standardizes wrist and hand pose representation across input devices."},{"cited_title":"Leap hand: Low-cost, efficient, and anthropomorphic hand for robot learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the LEAP hand, the 16-DoF anthropomorphic end-effector used in the main experiments."}],"review_version":1}