{"id":"45205f91-6d13-430e-8970-be2b3d9253f8","arxiv_id":"2506.08931","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"CLONE combines a mixture-of-experts teleoperation policy with closed-loop LiDAR odometry feedback to keep a humanoid's global position within about 5 cm while walking nearly 9 m.","lead":"CLONE is a humanoid teleoperation system that turns only head and hand tracking from a mixed-reality headset into coordinated whole-body robot motion, and uses LiDAR odometry to correct positional drift during long walks. In tests on a Unitree G1 robot, the system kept average position error around 5 cm over paths up to 8.9 m and demonstrated tasks such as walking, squatting, and jumping.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unvalidated AVP odometry reference: the 5.1 cm drift claim is measured relative to the same uncharacterized reference the closed loop chases.","rationale":"The reader's weakest assumption—that AVP odometry is an accurate, drift-free operator reference—is exactly the load-bearing point. I independently reached the same conclusion by tracing the real-world metric: the straight-path error is the robot position minus the AVP position (Sec. 4), the policy consumes that difference as the control target (Sec. 3.2), and the reward optimizes it (Table A1). No independent ground truth appears anywhere in the evaluation. I considered and set aside other candidate concerns: the missing same-protocol baseline weakens the 'unprecedented fidelity' wording but does not invalidate the drift experiment; the in-distribution simulation metrics are circular for generalization but the real-world demos are the primary support; and the internal editing note is a quality issue, not a correctness issue. The AVP-reference concern is more fundamental because it attacks the validity of the headline number itself. The proposed test is decisive and feasible: adding an external measurement to the existing protocol, or measuring AVP drift in isolation, would show whether the 5.1 cm error reflects true global tracking or just agreement between two drifting odometry systems. Since the reader's CONDITIONAL verdict already flags this assumption and requires additional evidence, my stress-test does not change the verdict.","tokens_in":17766,"tokens_out":5924,"duration_ms":61805,"concrete_test":"Repeat the straight-path experiment with an independent ground-truth measurement of the operator's true final position for each distance (3 m, 6 m, 8.9 m), using a laser rangefinder, total station, or external motion capture. Log three per-trial quantities: AVP-vs-true drift of the operator's headset, robot-vs-true error, and robot-vs-AVP error. If AVP-vs-true mean drift over 8.9 m is small (e.g., under 2 cm) and robot-vs-true error remains near the reported 5.1 cm mean / 12.0 cm max, the concern is resolved. If AVP drift is comparable to or larger than the reported errors, the headline 'minimal drift' claim does not establish global positioning accuracy, since the published metric is robot-vs-AVP only.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim—5.1 cm mean straight-path tracking error and 'minimal positional drift' over 8.9 m—is computed as the discrepancy between the robot's LiDAR-estimated final position and the operator's Apple Vision Pro (AVP) odometry position (Sec. 4; App. A.2). The closed-loop policy is explicitly trained to drive the robot to the AVP position: the task observation includes the AVP position and its difference from the robot's position (Sec. 3.2), and the reward minimizes that difference (Table A1, 'Body position (MR)'). Thus the reported error measures how well the robot tracks AVP's own estimate, not how accurately the robot tracks the operator's true global position. The paper never characterizes AVP odometry drift over 3–8.9 m trajectories, nor the frame alignment between AVP and LiDAR coordinates beyond 'fix[ing] the initial positions' (Sec. 4; Sec. 3.2). If AVP drifts, the robot will faithfully follow a drifting target, and the closed loop converts robot drift into reference drift; the 5.1 cm figure could then be an artifact of comparing two drifting estimators, not evidence of global positioning accuracy. The training noise model (Eq. 2) only adds a speed-dependent Wiener process to the head position; the constants cvel and cmin are not reported in Table A2, and the model does not capture slow VIO drift or frame misalignment. This concern is load-bearing because the abstract's 'minimal positional drift over long-range trajectories' and 'accurate global positioning' rest entirely on this measurement.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents CLONE, a teleoperation system for a Unitree G1 humanoid that uses only head and hand tracking from an Apple Vision Pro headset. The system trains a teacher policy in simulation with privileged state information and distills it into a student MoE policy that observes robot state history and task observations, including the difference between the robot's LiDAR odometry position and the operator's MR odometry position. The authors report real-world straight-path tracking with a mean error of 5.1 cm over 8.9 m, curved-path tracking with 20 cm mean error over 10 m, whole-body skill demonstrations, and simulation comparisons against an MLP baseline and a dataset ablation. The central claim is that closed-loop error correction prevents drift accumulation in long-horizon whole-body teleoperation while preserving coordination.","tokens_in":18053,"tokens_out":3886,"duration_ms":43235,"significance":"If the central drift claim is properly supported, CLONE would be a practically valuable contribution: it uses minimal commercial hardware, demonstrates real-world whole-body skills including squatting and jumping on a physical humanoid, and the closed-loop formulation directly addresses a known weakness of open-loop teleoperation. The MoE architecture and the CLONED dataset with hand-orientation augmentation are plausible contributions. However, the headline quantitative claim currently depends on an uncharacterized external reference and lacks a same-protocol baseline, so the significance for the community cannot yet be assessed at the level claimed in the abstract.","major_comments":[{"comment":"The 5.1 cm straight-path error is the discrepancy between the robot's LiDAR-estimated position and the operator's Apple Vision Pro odometry position, and this same AVP position is the reference signal that the closed-loop policy is trained to chase (Section 3.2, Table A1 'Body position (MR)'). The paper gives no characterization of AVP odometry drift over 3-8.9 m trajectories and no description of the calibration or frame-alignment procedure between the AVP frame and the LiDAR map frame beyond fixing initial positions. If the operator reference drifts, the robot will faithfully track a drifting target, and the reported error measures agreement between two drifting estimators rather than global positioning accuracy. This is load-bearing for the abstract claim of 'minimal positional drift over long-range trajectories'; the authors should report reference drift on the same paths or a stationary AVP test, and describe the frame-alignment procedure.","section":"§4, App. A.2"},{"comment":"The improvement over existing open-loop systems is asserted but not measured. The real-world path-following experiment has no baseline run under the same protocol, e.g., the same policy with the closed-loop error-correction term disabled, or an existing open-loop teleoperation system such as OmniH2O on the same 3/6/8.9 m straight path. Without a same-protocol baseline, the 5.1 cm figure is an absolute performance number; it does not support 'improved tracking precision over existing open-loop approaches' or 'unprecedented' fidelity. Adding an open-loop control condition in the same environment is necessary to substantiate the central contribution.","section":"§4"},{"comment":"The simulation metrics in Table 1 are computed on CLONED, the same dataset used to train the policy, and all methods report 100% success rate, so these numbers quantify in-distribution tracking rather than generalization. Figure 7 uses height-edited sequences from CLONED, which is a mild distribution shift but still within the source motion repertoire. The paper should state the train/test separation explicitly, or evaluate on held-out motion sequences and unseen stance configurations, to support the claim of robust generalization.","section":"Table 1, Fig. 7"},{"comment":"The velocity-dependent SDE noise model is presented as reflecting real LiDAR and IMU error characteristics, but the constants cvel and cmin are not reported in Table A2, and the model adds noise only to the head position, not to the operator's global reference. Slow VIO drift and frame misalignment between operator and robot odometry are precisely the error modes that the closed-loop claim depends on, yet they are absent from the training noise model. Please report the constants and either justify the omission of reference drift in training or include a variant that models it.","section":"Eq. (2), Table A2"}],"minor_comments":[{"comment":"The figure label 'Close-Loop Tacking' contains typos; it should read 'Closed-Loop Tracking'.","section":"Fig. 2"},{"comment":"The word 'teloperation' is misspelled; it should be 'teleoperation'.","section":"App. A.1"},{"comment":"The plot shows mean errors but no error bars or per-condition standard deviations, although each straight-path condition was repeated ten times; reporting spread would help assess consistency.","section":"§4, Fig. 4"},{"comment":"The subsection title 'Qualitative Results Comparsion' should be 'Qualitative Results Comparison'.","section":"App. D.3"},{"comment":"In the legend, 'CLONE †' and 'CLONE *' are rendered with inconsistent spacing; this is cosmetic but should be fixed for clarity.","section":"§5, Fig. 7"},{"comment":"The t-tests comparing distance groups have only n=10 per group; the absence of a significant difference is weak evidence of no drift accumulation and should be phrased accordingly.","section":"§4"}],"recommendation":"major_revision","confidential_remarks":"The core idea is timely and the real-hardware results are encouraging, but the headline drift number is currently defined relative to the same unvalidated reference that the controller chases. The requested additions—AVP drift characterization, a same-protocol open-loop baseline, and a clearer train/test split for simulation—are within reach and would turn this into a solid contribution. I do not see a reason to reject at this stage, but the manuscript is not ready as is."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"CLONE is a solid systems integration paper, not a breakthrough. What is genuinely new is the package: an MoE student policy, LiDAR odometry feedback folded into the task observation, and a hand-orientation-augmented dataset, all running on a real Unitree G1 with commodity MR input. That combination is practically useful for humanoid teleoperation and data collection, and the real-robot demos—squatting, jumping, the 15 m mixed navigation—are credible evidence that the closed loop works at a systems level. The ablations against an MLP student and OmniH2O-style training data are meaningful, and the expert-activation analysis gives a real look at how the MoE specializes. The authors also cite the prior MoE and legged-odometry literature instead of hiding it. Credit where due: this is a working system, and the integration is nontrivial.\n\nThe soft spots are concentrated in the headline quantitative claim. The 5.1 cm mean error over 8.9 m is the discrepancy between the robot's LiDAR-estimated position and the operator's Apple Vision Pro position. The policy is explicitly trained to minimize exactly that difference, and the reward table confirms it. So the number measures how well the robot tracks AVP's own estimate, not how accurately the robot knows its global position. The paper never characterizes AVP odometry drift over these trajectories, and the frame alignment between AVP and LiDAR is described only as 'fixing the initial positions.' If AVP drifts slowly, the robot chases a moving target, and the reported error can stay small while both estimates are wrong. That gap is load-bearing for the 'unprecedented fidelity' and 'minimal positional drift' language in the abstract.\n\nThe other weaknesses are more moderate. There is no same-protocol open-loop or prior-system baseline, so the improvement claim over open-loop teleoperation is asserted rather than demonstrated. Table 1 is computed on CLONED, the training distribution, so those numbers measure in-distribution tracking; the stance-height experiments are edited from the same data, so generalization evidence is limited. No code, data, or trained models are released, which hurts reproducibility for a paper that is half dataset and architecture. There is also a leftover Chinese editing note in the Figure 2 caption that should be cleaned before any archival version.\n\nOverall, the engineering story is plausible and the real-robot feasibility is credible, but the central global-accuracy number needs an independent ground-truth baseline before it supports the abstract's claims. This paper deserves a serious referee: the system is worth engaging with, and a good review could force the missing measurement and make the paper much stronger. Send it to review, with the clear instruction that the AVP-reference problem must be handled head-on.","headline":"Solid closed-loop teleoperation integration, but the 5.1 cm drift number is measured against the same AVP reference the loop chases, so the global-accuracy claim outruns the evidence.","tokens_in":18671,"tokens_out":2298,"would_cite":false,"duration_ms":23626,"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":"A closed-loop teleoperation system keeps a humanoid within centimeters of an operator over long trajectories using only head and hand tracking from an MR headset.","keywords":["humanoid teleoperation","whole-body control","mixture-of-experts","closed-loop error correction","LiDAR-inertial odometry","long-horizon tasks","loco-manipulation","motion dataset"],"falsifier":"Repeat the straight-path experiment while independently tracking both the operator and the robot with an external motion-capture or surveyed ground truth; if the robot's absolute endpoint error grows with distance while the operator-relative error stays small, the bounded drift is real, whereas if both errors grow together, the closed loop has only synchronized the two odometry drifts.","tokens_in":17488,"feed_emoji":"🤖","tokens_out":9154,"duration_ms":91890,"temperature":0.7,"pith_summary":"CLONE is a humanoid teleoperation system that claims to solve two problems that have limited prior work: whole-body coordination and unbounded positional drift. It uses only the poses of both hands (position and orientation) and the position of the head from a commercial mixed-reality headset, and it learns a single Mixture-of-Experts policy that turns those sparse signals into joint commands for the whole body, including locomotion and squatting. Drift is addressed by closing the loop: LiDAR-inertial odometry gives the robot's global position, the headset gives the operator's global position, and the policy is trained to minimize the difference between the two. In real-world tests on a physical humanoid, the paper reports a mean straight-path tracking error of 5.1 cm up to 8.9 m, a maximum deviation of 12.0 cm, a curved-path mean error of 20 cm with 2 degrees of rotational drift over a 10 m trajectory with two 90-degree turns, and a 15 m mixed sequence that returns near its start. If these results hold, long-horizon teleoperation no longer needs full-body motion capture or open-loop control shortcuts.","feed_headline":"Humanoid follows operator 8.9 m with only 5.1 cm drift","feed_subtitle":"Closed-loop LiDAR feedback plus a mixture-of-experts policy keep long-range teleoperation on track using just head and hand tracking.","key_machinery":"The load-bearing mechanism is a closed-loop global-position error signal: the robot's position $p \\in \\mathbb{R}^3$ comes from LiDAR-inertial odometry at 10 Hz, the operator's position $\\hat{p} \\in \\mathbb{R}^3$ comes from the MR headset's odometry, and the policy's task observation includes the difference $\\hat{p} - p$, so actions are chosen to reduce it. Around that signal, the student policy is a Mixture-of-Experts network with $L=3$ layers and $N=4$ experts per layer, taking the top $k=2$ experts by routing weight, with a balancing loss that keeps expert usage uniform; it is trained by distillation from a teacher policy and regularized by an adversarial motion prior that shapes lower-body behavior. A velocity-dependent stochastic noise model on the head position is added during training so the policy learns to tolerate odometry error that grows with movement speed.","core_discovery":"On its own terms, the paper establishes that a teleoperation policy can preserve accurate global positioning over long trajectories while keeping whole-body coordination, using only sparse upper-body references. The core discovery is that including the closed-loop difference between the humanoid's LiDAR-estimated global position and the operator's headset-estimated global position in the policy observation changes the failure mode of teleoperation: instead of stepwise tracking error accumulating, the robot continuously corrects toward the operator, and the trained policy turns that correction into natural whole-body motion. The same policy, through routed expert specialization, reproduces standing, walking, turning, side-stepping, squatting, and jumping, with experts splitting along posture lines and all experts engaging during dynamic motions. The paper further claims that its curated CLONED dataset, though far smaller than prior motion-capture collections, outperforms the larger dataset on crouch and squat-to-pick tasks, and that the MoE architecture beats a monolithic network on local motion metrics and hand orientation tracking.","pith_inferences":["The same closed-loop recipe transfers to other leader-follower pairs with independent odometry, so a different robot or a simulated avatar could inherit the drift bound without re-deriving the correction logic.","Performance should degrade in environments where the headset's visual-inertial odometry degrades, such as low light, sparse texture, or fast head motion, because the error signal itself becomes noisy.","If the operator reference drifts, the system converts unbounded robot drift into bounded operator-relative drift, so the headline error numbers should be read as error with respect to the operator, not with respect to the world.","The data-efficiency result suggests a cost-effective recipe for other whole-body control tasks: a few hundred curated, motion-edited sequences may beat much larger unfiltered collections."],"forward_implications":["In the tested distance range, tracking error stops growing with distance: statistical tests across 3 m, 6 m, and 8.9 m show no significant degradation, so errors no longer compound per step.","Whole-body coordination, including squatting and standing from a squat, can be generated from sparse input alone, with lower-body behavior shaped by the adversarial motion prior rather than explicit lower-body commands.","A single policy can cover diverse skills without mode collapse when the student network uses routed experts rather than a monolithic network.","Dataset composition, including hand orientations and transition sequences, can matter more than raw dataset size for reproducing manipulation-relevant postures.","LiDAR-inertial odometry running at 10 Hz is sufficient to bound position error when the policy runs at 50 Hz and consumes the latest odometry reading."],"supporting_citations":[{"why":"Supplies the LiDAR-inertial odometry that gives the robot its global position for the closed-loop correction.","marker":"[27]"},{"why":"Provides the teacher-student training framework, reward terms, and domain randomization that CLONE builds on and compares against as the baseline.","marker":"[18]"},{"why":"AMASS is the base motion-capture dataset that CLONED filters, edits, and augments with hand orientations and custom sequences.","marker":"[28]"},{"why":"The adversarial motion prior supplies the reward that keeps lower-body motion natural when only upper-body references are given.","marker":"[66]"}],"fun_headline_variants":["Closed-loop humanoid teleop keeps drift to 5 cm over 9 m","MoE + LiDAR feedback make humanoid follow operator precisely","Sparse reference, 5 cm drift: whole-body teleop for long horizons","CLONE teleop: closed-loop, whole-body, low drift over long tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The closed-loop controller treats the operator's global position, as estimated by the MR headset's motion tracking, as the true reference the robot should match, so any drift in that headset estimate becomes robot error.","fun_headline_variants_meta":{"raw":{"variants":["Closed-loop humanoid teleop keeps drift to 5 cm over 9 m","MoE + LiDAR feedback make humanoid follow operator precisely","Sparse reference, 5 cm drift: whole-body teleop for long horizons","CLONE teleop: closed-loop, whole-body, low drift over long tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001721,"raw_usage":{"total_tokens":6803,"prompt_tokens":934,"completion_tokens":5869,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":5786}},"tokens_in":550,"tokens_out":5869,"duration_ms":45366,"temperature":1.0,"reasoning_tokens":5786,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:58:52.984613+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the straight-path experiment while independently tracking both the operator and the robot with an external motion-capture or surveyed ground truth; if the robot's absolute endpoint error grows with distance while the operator-relative error stays small, the bounded drift is real, whereas if both errors grow together, the closed loop has only synchronized the two odometry drifts.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the LiDAR-inertial odometry that gives the robot its global position for the closed-loop correction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the teacher-student training framework, reward terms, and domain randomization that CLONE builds on and compares against as the baseline."},{"cited_title":"Mahmood, N","cited_arxiv_id":null,"evidence_quote":"AMASS is the base motion-capture dataset that CLONED filters, edits, and augments with hand orientations and custom sequences."},{"cited_title":"crouch” or “squat to pick up an object","cited_arxiv_id":null,"evidence_quote":"The adversarial motion prior supplies the reward that keeps lower-body motion natural when only upper-body references are given."}],"review_version":1}