{"id":"b82984c2-6151-4d5a-afe3-7d6c6513e482","arxiv_id":"2412.20226","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Training a diffusion-based navigation policy on human-held depth camera demonstrations lets a quadruped robot reach and follow a human target while avoiding obstacles.","lead":"Researchers trained a robot navigation policy from 1.5 hours of demonstrations recorded by a person walking with a handheld depth camera, with no robot involved in data collection. The policy then guided a quadruped robot to reach and follow a human target while avoiding obstacles, outperforming comparison methods in the reported tests.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Embodiment-agnostic claim is under-tested: only scaler velocity rescaling adapts the policy across platforms, with evidence from a single robot.","rationale":"The reader's weakest assumption identifies exactly the load-bearing concern: the transfer from human-held camera demonstrations to a robot relies on a minimal linear normalization and is tested on only one embodiment. This is the central claim of the paper, as the title and Figure 1 assert that the policy can be deployed on 'any robot.' The empirical evidence supports that the transfer works for the Unitree Go2, but it does not support the general 'embodiment-agnostic' claim. The paper's strengths include real-robot experiments, open-source code, and a low-data regime, but these do not compensate for the absence of multi-embodiment evaluation. I considered the unfair baseline modification (ViNT with target image replaced by relative distance) as an alternative concern, but it is secondary: even if the baselines were perfectly fair, the embodiment-agnostic claim would still be unsupported. The proposed concrete test—varying camera height on the same robot—directly probes the most likely failure mode without requiring a second robot platform. Therefore, the reader's conditional verdict is appropriate; no verdict change is needed.","tokens_in":9776,"tokens_out":7084,"duration_ms":72395,"concrete_test":"Mount the same depth camera on the Unitree Go2 at a substantially different height (e.g., 1.0–1.2 m, typical of wheeled platforms) while keeping the ViDEN policy unchanged and using only the max-velocity denormalization from Section II-B. Run the same 25-trial hard-level protocol from Section III-C. If the success rate drops markedly (e.g., from 84% to below 50%), the policy is sensitive to sensing embodiment, contradicting the 'embodiment-agnostic' claim. A complementary simulation test with a unicycle model of different max speed and turning radius would isolate the effect of dynamic differences on the transferred trajectories.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central novelty of ViDEN is that a policy trained from human-held camera demonstrations, with no robot in the loop, can be deployed on arbitrary embodiments (title, Fig. 1, Section II-B). The only embodiment adaptation described is a scalar velocity normalization/denormalization (Eq. 1, Section II-B), which rescales waypoints by the robot's maximum velocity. No account is taken of differences in kinematics (turning radius, acceleration limits), dynamics, camera height, field of view, or control bandwidth. The paper evaluates deployment on exactly one robot (Unitree Go2, Section III-B), with a camera height comparable to the 0.3–0.5 m used during data collection. This single positive transfer is insufficient to support the 'embodiment-agnostic' claim, which is the load-bearing premise of the title and the framework's stated generality. Moreover, Eq. (1) as written normalizes a position by quantities called 'motion velocities,' which is dimensionally inconsistent; if the actual implementation uses displacement or velocity scaling, the description is imprecise, and if it truly uses only max velocity, it is too simplistic to capture embodiment differences. The paper's own generalization experiment (Section III-C) tests a new environment, not a new embodiment, so the embodiment-agnostic aspect remains untested beyond the Go2.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ViDEN, a behavior-cloning framework that trains a diffusion-based navigation policy entirely from human demonstrations recorded with a hand-held depth camera. The policy consumes depth images plus a relative target position and outputs a short waypoint trajectory, which is then re-scaled to the robot's maximum velocity. The authors claim that this makes the policy embodiment-agnostic. The main experiments deploy the policy on a Unitree Go2 quadruped for human reaching/tracking in easy, moderate, and hard tasks, and compare against a BC-ConvMLP baseline and a modified ViNT baseline. Additional experiments measure robustness to low light, physical perturbations, and dynamic obstacles, as well as generalization to a new indoor environment with varying amounts of fine-tuning data. The central claim is that ViDEN achieves higher success rates than the baselines with only about one hour of robot-independent data.","tokens_in":10077,"tokens_out":4335,"duration_ms":45542,"significance":"If the empirical results hold, the paper makes a useful contribution to demonstration-based navigation: it shows that a diffusion policy trained on data collected without any robot can be deployed on a legged platform, and that depth plus relative target conditioning can reduce the data requirement compared to image-goal methods such as ViNT. The open-source code and the emphasis on low-cost data collection are concrete strengths. However, the significance is currently limited by three issues: the 'embodiment-agnostic' claim is supported by a single robot and a scalar re-scaling; the success-rate evaluation lacks an explicit success criterion, error bars, and statistical detail; and the ViNT baseline is modified in a way that may not be a faithful comparison. These issues are addressable, but they are load-bearing for the paper's central claims.","major_comments":[{"comment":"The embodiment-agnostic claim is the load-bearing premise of the title and Section II-B, but it rests on a single experiment on a Unitree Go2. The only embodiment adaptation described is a scalar re-scaling (Eq. (1) and the 'denormalization with its maximum velocity' in Section II-E), which does not account for differences in kinematics, dynamics, camera height, field of view, or control bandwidth. As written, Eq. (1) also has a units problem: p_t is a position, while β_min and β_max are described as 'motion velocities,' so the normalized quantity is dimensionally inconsistent unless β are actually displacements or the text is rephrased. To support the title's claim, please test on at least one additional embodiment with significantly different scale or kinematics, or explicitly scope the claim to e.g., 'legged platforms with comparable camera geometry.'","section":"II-B, III-B"},{"comment":"The success rates are the paper's central quantitative evidence, but the manuscript never defines what counts as a success: there is no distance threshold to the human, no time limit, no collision tolerance, and no specification of whether the trial ends when the target is lost. With 25 trials per condition and 5 per scene, reporting only aggregate percentages without error bars or per-scene variability makes it impossible to judge whether the differences between ViDEN (96%, 84%) and its simplified version (88%, 72%) or between ViDEN and ViNT are meaningful. Please report the exact success definition, the number of trials per cell, confidence intervals, and preferably per-scene results.","section":"III-C, Table I"},{"comment":"The ViNT baseline is modified by replacing its target image with the relative distance to the target. This is a substantive alteration to the model's conditioning mechanism, and ViNT is designed and trained for image goals, so the comparison may not reflect ViNT's intended performance. The manuscript should state exactly how the modification was implemented: which modules were replaced, what was retrained, and with what data and hyperparameters. Without this, the claim that ViDEN outperforms 'existing methods' is not fully supported.","section":"III-C, baselines"},{"comment":"The reported data volume is inconsistent: Section I states '1.5 hours of robot-independent data,' while Section III-A reports 120 demonstrations with an average duration of 30 seconds, which is about 60 minutes. In addition, several critical hyperparameters (λ=2.5 m set by 'trial and error,' the β scaling, goal masking probability pm=0.7, task-agnostic fraction κ=20%, and target dropout 50%) are tuned for the evaluation environment without any sensitivity analysis. Since these choices directly influence the success rates in Table I and Figure 7, please clarify the actual data budget and provide at least a minimal sensitivity study or explicitly state that the main results are not sensitive to these values.","section":"III-A, III-E"}],"minor_comments":[{"comment":"Equation (1) and the surrounding text should use consistent units and explicitly state whether the normalized quantity is a displacement, velocity, or pose waypoint; the current phrasing 'position' versus 'motion velocity' is confusing.","section":"II-B"},{"comment":"The data amount should be aligned: Section I says 1.5 hours, while Section III-A says approximately 60 minutes; please make the numbers consistent and clarify whether task-agnostic data is additional.","section":"III-A"},{"comment":"The robustness table reports success rates but does not specify the number of trials per disturbance type or the magnitude and duration of the perturbations; please add this information so the rates are interpretable.","section":"III-C, Table II"},{"comment":"The claim of outperforming 'existing methods' should be qualified: the only named baseline from the recent literature is ViNT, and it is used in modified form; GNM and NoMaD are discussed but not evaluated.","section":"III-C"},{"comment":"The generalization plot has no error bars or trial counts; given that the zero-shot success rate is 55%, please provide confidence intervals to support the conclusion that zero-shot transfer is feasible.","section":"III-C, Figure 7"}],"recommendation":"major_revision","confidential_remarks":"The core idea is attractive and the real-robot experiments are a valuable contribution, but the current manuscript overstates the embodiment-agnostic claim relative to the evidence. The main technical concerns—success criteria, error bars, baseline fidelity, and the units of Eq. (1)—are fixable in revision. I would encourage the editor to request a revision that either adds a second robot or re-scopes the claim, and that tightens the empirical reporting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The useful core here: they train a diffusion policy on about 1.5 hours of demonstrations collected by a human walking a handheld depth camera, no robot in the loop, then deploy it on a Unitree Go2 quadruped and get 84% success on their hard following task. That is a real result and a genuinely useful data-collection shortcut. Unlike GNM/ViNT/NoMaD, which need tens of hours of robot-collected data across platforms, this gets a working policy from a human wandering around. As a data-collection method, it's a reasonable contribution.\n\nWhat's good: the real-robot testing is a strength. They compare against BC-ConvMLP and ViNT, and they do ablations that isolate the effects of target labels, goal masking, and task-agnostic data. The robustness results honestly show weak spots (physical perturbations at 30% success), and the zero-shot generalization to a new environment is 55%, reported without spin.\n\nNow the soft spots. The title says 'embodiment-agnostic' and the paper claims transfer to arbitrary robots. The evidence is one quadruped with a camera height similar to data collection, and the only cross-embodiment adaptation is a scalar scaling of output waypoints by the robot's max velocity. That's a reasonable first step, but it does not support the scope of the claim. The generalization experiment is a new environment, not a new embodiment. Also, Eq. (1) normalizes a position by quantities called 'motion velocities,' which is dimensionally inconsistent on the page. Either the implementation is doing something else, like displacement scaling, or the description is imprecise. A referee should ask for clarification and ideally a second robot.\n\nThe evaluation is also thinner than it looks. Success rates come from 25 trials per condition but there are no error bars or confidence intervals, so differences like 92% vs 96% are not meaningful without variance. The ViNT baseline is modified to use relative distance instead of a goal image, which could handicap it; that's worth noting. Hyperparameters like lambda and the masking probability are tuned on the Go2, so they are not embodiment-agnostic.\n\nNone of this kills the central result: human-held camera data can train a navigation policy that works on a legged robot. That is worth having. But the broad claim is not yet supported, and the evaluation needs more rigor.\n\nWho should read it: people building data collection pipelines for navigation LfD, and anyone working on target following from egocentric depth. It deserves a serious referee. The method is reproducible in principle, the experiments are real, and the limitations are honestly stated. I'd lean accept with major revisions: tone down the 'embodiment-agnostic' title claim, add statistical analysis, and clarify the normalization. Candidly, it's a good paper, not a great one, and a solid referee would make it better.","headline":"A useful data-collection trick for navigation, wrapped in an overbroad 'embodiment-agnostic' claim; the real-robot results are real but the evaluation is thinner than the title suggests.","tokens_in":10550,"tokens_out":3354,"would_cite":true,"duration_ms":33232,"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":"The paper claims that a diffusion-based navigation policy trained on 1.5 hours of human-held depth-camera demonstrations, with no robot in the loop, can be deployed on a legged robot to reach and track a human while avoiding obstacles.","keywords":["depth-based navigation","diffusion policy","behavior cloning","embodiment-agnostic","human demonstration","target tracking","legged robot","visual navigation"],"falsifier":"Run the same trained policy on a second embodiment, such as a wheeled robot or a taller legged robot, in the same easy, moderate, and hard tasks. If success rates drop far below the reported 100%, 96%, and 84%, the embodiment-agnostic transfer claim fails.","tokens_in":9604,"feed_emoji":"🤖","tokens_out":5396,"duration_ms":50719,"temperature":0.7,"pith_summary":"ViDEN claims that a robot navigation policy can be trained entirely from demonstrations recorded by a person walking a handheld depth camera, with no robot in the loop, and then deployed on a different embodiment. The paper argues that representing actions as normalized waypoints in the camera frame, conditioning on the target's relative position, and generating trajectories with a diffusion model makes the transfer possible. On a Unitree Go2 quadruped, the resulting policy reaches and tracks a human while avoiding obstacles, with success rates of 100%, 96%, and 84% on easy, moderate, and hard tasks using only about 1.5 hours of collected data. This matters because demonstration-based navigation is usually tied to a specific robot or needs 100+ hours of robot-collected data, while ViDEN offers a cheaper, robot-independent route to goal-directed navigation. The same framework also shows zero-shot transfer to a new indoor environment at 55% success, improving with modest fine-tuning data.","feed_headline":"Human-walked camera demos train a legged robot to navigate","feed_subtitle":"A depth-image diffusion policy reaches and tracks humans around obstacles, beating stronger baselines with far less data.","key_machinery":"The load-bearing mechanism is the combination of an embodiment-agnostic action space and a goal-conditioned diffusion policy. Actions are trajectories of camera-frame waypoints normalized by Eq. (1), so the policy outputs the same normalized relative motion regardless of robot; deployment rescales by the robot's maximum velocity. Depth frames are encoded with a CNN, target position and a randomly sampled intermediate goal are encoded with MLPs, and the transformer-encoded latent is fed to a U-Net diffusion model that denoises an action trajectory over K iterations. The paper's contribution is that this pipeline transfers from a human-held camera at 0.3–0.5 m height to a quadruped, because the policy learns task-centric geometry rather than robot-specific dynamics.","core_discovery":"On the paper's own terms, the discovery is that depth images plus relative target position provide enough information for a diffusion-based behavior-cloning policy to synthesize collision-free, target-directed trajectories, and that the demonstration data need not come from the robot at all. A human collects 120 trajectories by walking with a hand-held stereo camera toward a person while passing obstacles; the target is extracted with a detector, the camera trajectory is recorded as $SE(2)$ poses, and the policy is trained to output 32-step normalized waypoint trajectories conditioned on depth and on an intermediate goal. When deployed on a quadruped with a PD controller, the policy completes reaching and tracking tasks at higher success rates than a convolutional MLP baseline and than a goal-conditioned ViNT variant adapted to relative targets, across easy, moderate, and hard obstacle configurations. The paper further reports that task-agnostic exploration data, goal masking, dropout of target position, and a memory token each contribute to robustness, and that the policy can partially infer targets from depth without explicit labels.","pith_inferences":["The same data-collection scheme should extend to other target classes, such as animals, vehicles, or arbitrary objects, and to other robots with different camera placements, but only if viewpoint differences are handled; testing on a second embodiment would clarify the limits.","A natural next step the paper leaves implicit is to split the learned policy into a high-level goal and trajectory planner and a per-robot low-level controller, which would likely improve robustness to physical perturbations, where the reported success drops to 30%.","Because the policy sees depth only, its success in low light (73%) may reflect the depth sensor's resilience; an RGB-only variant would isolate whether depth, rather than the diffusion architecture, drives the transfer."],"forward_implications":["Navigation policies can be trained without any robot in the loop, so a new target-following capability can be obtained with about 1.5 hours of human walking instead of 100+ hours of robot teleoperation.","Because goals are specified by relative position rather than goal images, the policy can follow a moving human and does not require a topological map or pre-specified target image.","Goal masking lets a single policy switch between directed navigation and exploration, giving it some ability to infer targets from depth alone when explicit target labels are absent.","Deploying in a new environment needs only about 50% of the demonstration data of the original environment to match training performance, suggesting rapid adaptation to new scenes."],"supporting_citations":[{"why":"introduces the normalized waypoint action space that ViDEN adapts for embodiment transfer.","marker":"[25]"},{"why":"provides the ViNT baseline that the paper modifies to use relative target distance instead of goal images.","marker":"[26]"},{"why":"supplies the goal-masked diffusion policy formulation that ViDEN builds on for conditioned trajectory generation.","marker":"[27]"},{"why":"gives the diffusion-policy training and denoising machinery used to output action trajectories.","marker":"[28]"},{"why":"shows in-the-wild human-collected demonstration data can train policies without robot hardware, supporting the data-collection premise.","marker":"[20]"},{"why":"demonstrates human-held data collection for mobile robot policies, a precedent for robot-free demonstration acquisition.","marker":"[30]"}],"fun_headline_variants":["Human walk demos train robot navigation with depth","Depth diffusion policy learns navigation from human demos","Robot nav trained on human camera walks needs little data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that demonstrations recorded by a person walking with a handheld camera at 0.3–0.5 m height can be executed by a different robot after linear velocity normalization, without modeling differences in kinematics, dynamics, sensor height, or field of view, and this premise is tested on only one quadruped.","fun_headline_variants_meta":{"raw":{"variants":["Human walk demos train robot navigation with depth","Depth diffusion policy learns navigation from human demos","Robot nav trained on human camera walks needs little data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000256,"raw_usage":{"total_tokens":1565,"prompt_tokens":927,"completion_tokens":638,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":543,"completion_tokens_details":{"reasoning_tokens":590}},"tokens_in":543,"tokens_out":638,"duration_ms":7373,"temperature":1.0,"reasoning_tokens":590,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:25:27.522445+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same trained policy on a second embodiment, such as a wheeled robot or a taller legged robot, in the same easy, moderate, and hard tasks. If success rates drop far below the reported 100%, 96%, and 84%, the embodiment-agnostic transfer claim fails.","supporting_citations":[{"cited_title":"GNM: A general navigation model to drive any robot,","cited_arxiv_id":null,"evidence_quote":"introduces the normalized waypoint action space that ViDEN adapts for embodiment transfer."},{"cited_title":"ViNT: A foundation model for visual navigation,","cited_arxiv_id":null,"evidence_quote":"provides the ViNT baseline that the paper modifies to use relative target distance instead of goal images."},{"cited_title":"NoMaD: Goal masked diffusion policies for navigation and exploration,","cited_arxiv_id":null,"evidence_quote":"supplies the goal-masked diffusion policy formulation that ViDEN builds on for conditioned trajectory generation."},{"cited_title":"Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots,","cited_arxiv_id":null,"evidence_quote":"shows in-the-wild human-collected demonstration data can train policies without robot hardware, supporting the data-collection premise."},{"cited_title":"UMI on legs: Making manipulation policies mobile with manipulation-centric whole-body controllers,","cited_arxiv_id":null,"evidence_quote":"demonstrates human-held data collection for mobile robot policies, a precedent for robot-free demonstration acquisition."}],"review_version":1}