{"id":"81b345a4-97df-4155-93ac-a89dbb64d93d","arxiv_id":"2610.02840","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Forecasting scene and hand motion together as 3D point trajectories, with pre-training on human videos, lifts DexJoCo success to 69.0%, 11.7 points above the strongest baseline.","lead":"PointWAM encodes the scene and the robot's hands as colored 3D points, forecasts how both move as point trajectories in one shared space-time frame, then converts the predicted hand motion into robot commands.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"The 10.9-point scene-supervision gain is ablated by removing supervision from both pre-training and fine-tuning simultaneously, so the gain cannot be attributed to the estimated human-video scene trajectories; it may come entirely from accurate simulator targets at fine-tuning.","rationale":"The reader's verdict (CONDITIONAL, high confidence) is well-founded: the headline success numbers are supported by clean within-method ablations and external baselines, and the paper is honest about its limitations. My read agrees with keeping the paper CONDITIONAL, but for a partly different reason than the reader's weakest_assumption. The reader located the risk in the unvalidated quality of the CoTracker3+VGGT-Omega scene estimates. I argue the sharper and more load-bearing gap is attribution: the scene-supervision ablation (Table 3b) removes the supervision from both pre-training and fine-tuning at once, and these stages use different target sources — estimated (human videos) versus accurate simulator replay (DexJoCo, §B.2). Even if the estimates were perfect, the 10.9-point gain could be entirely a fine-tuning auxiliary-loss effect, in which case the paper's claim that human-video scene forecasting powers world modeling is unsupported. Conversely, if a pre-training-only ablation reproduces the gain, the estimates become load-bearing and must be validated before the claim is trusted. Neither the reader's framing nor the paper's current experiments distinguish these. The concrete pre-training-only ablation is a single, inexpensive run that settles the question. Secondary concerns (the 2D-vs-3D ablation confounded with encoder choice in Table 3b, checkpoint selection on the first seed's test average in C.2, and the 24-trial real-robot evaluation without error bars) are all real but moderate; they would move the paper from clean to strong but do not touch the core success-rate result, whereas the scene-gain attribution determines whether the paper is a world-action model claim at all. I take no issue with the honesty of the write-up: the paper states its own limitations (b14, b164) and the non-decomposition is a design matter, not a misrepresentation. Conditional acceptance is the right call, with a concrete ablation addressing attribution as the path to accept.","tokens_in":25684,"tokens_out":7284,"duration_ms":51410,"concrete_test":"Run the reference ablation (Volt-L12 pre-trained on EgoDex, fine-tuned on DexJoCo) but remove scene-trajectory supervision from pre-training only, keeping it in fine-tuning: pre-train the backbone and hand head with hand-trajectory supervision alone, then fine-tune with the full L_traj of Eq. 7 as in the reference. Compare DexJoCo SR to the reference 64.4% and to the 53.5% no-scene-supervision row. If SR stays at ~64%, the 10.9-point gain comes from accurate fine-tuning scene targets and the estimated human-video scene supervision is not load-bearing, which would reframe the human-video world-modeling claim.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's world-modeling claim rests on Table 3(b) (wording in §4.4: 'Removing scene-trajectory supervision from both pre-training and fine-tuning'), which yields the 10.9-point figure quoted in the abstract (b5) and contributions (b25). The two stages use different, not interchangeable, scene targets. In fine-tuning on DexJoCo, scene trajectories come from replaying demonstrations with simulator segmentation, which assigns each point to an object or rigid part and defines its trajectory by rigid motion (§B.2); these targets are accurate. In pre-training on human videos, scene trajectories come from CoTracker3 2D tracks lifted with VGGT-Omega monocular depth, whose metric scale is recovered by assuming the ARKit/VITRA hand annotations are correct (§B.1 'Depth', 'Scene tracks'). The ablation removes the supervision in both stages at once, so it cannot distinguish the contribution of the accurate fine-tuning targets from the contribution of the estimated pre-training targets. Consequently the reader's weakest_assumption — that the load-bearing supervision for the world-modeling gain is the unvalidated human-video estimate — is untested in either direction: the 10.9 points could come entirely from the accurate fine-tuning scene loss (an auxiliary-task or regularization effect on simulator targets), with the estimated human-video scene trajectories contributing nothing, or even posing noise the model must unlearn in fine-tuning. If that were true, the paper's distinctive conclusion (b163) that a shared 3D point-trajectory representation lets robots learn dexterous manipulation 'largely from human videos' would weaken: the model would be a hand-keypoint retargeting policy whose scene forecast matters only on robot data. This is a sharper gap than merely 'the estimates are unvalidated': validating the estimates (the reader's suggested remedy) would not settle whether they are load-bearing, since the gain might live entirely in the accurate fine-tuning stage.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper introduces PointWAM, a world action model that decomposes the world into a scene and hands and forecasts both as 3D point trajectories in a shared space-time frame. Given a colored point cloud and a language instruction, a 3D trajectory forecaster predicts per-point scene displacements and per-keypoint hand displacements over a T=30 horizon, and an action retargeter converts the forecast hand motion into robot action chunks. The model is pre-trained on 1.15M EgoDex/VITRA human demonstration clips (hand keypoints from annotations, scene tracks from CoTracker3 tracks lifted with VGGT-Omega depth scaled by the hand annotations), fine-tuned on DexJoCo demonstrations (scene trajectories from simulator replay with rigid-part segmentation), and evaluated on ten DexJoCo tasks, eight RoboDojo-Precision tasks, and two real-robot tasks. Headline results are 69.0% average multi-task success on DexJoCo (+11.7 percentage points over GR00T N1.6), +56.9 points from human-video pre-training, +10.9 points from scene-trajectory supervision (Table 3b), and consistent real-robot gains over GR00T N1.6 and π0.5.","tokens_in":26028,"tokens_out":10573,"duration_ms":68032,"significance":"If the reported results hold, PointWAM makes a strong empirical case that explicit 3D point-trajectory world modeling, pre-trained on human egocentric video, can outperform robot-pre-trained VLAs on dexterous manipulation both in simulation and on a small real-robot evaluation. The paper deserves credit for unusually transparent experimental hygiene: the pre-training gain (Fig. 6) and the encoder comparison (Table 3a) hold the rest of the pipeline fixed, training settings are tabulated (Table 4), baseline protocols are specified in detail (App. C.5), and acknowledged limitations (within-clip point correspondence, point-spacing resolution) are stated. The load-bearing gap is the attribution of the scene-supervision benefit: the single ablation removes scene supervision from both training stages at once, so the reported 10.9-point gain is a whole-system property that does not, by itself, identify the estimated human-video scene trajectories as the source of the gain.","major_comments":[{"comment":"The 10.9-point scene-supervision gain, quoted in the abstract (b5), the contribution list (b25), and the conclusion (b163), is measured by removing scene-trajectory supervision from pre-training and fine-tuning simultaneously ('Removing scene-trajectory supervision from both pre-training and fine-tuning'; App. C.1: 'Every variant ... changes both pre-training and fine-tuning'). The two stages use scene targets that are not interchangeable: at fine-tuning on DexJoCo the targets come from simulator replay with rigid-part segmentation and are accurate (App. B.2), whereas in pre-training on human videos they come from CoTracker3 2D tracks lifted with VGGT-Omega monocular depth whose metric scale is recovered by assuming the ARKit/VITRA hand annotations are correct (App. B.1, 'Depth' and 'Scene tracks'). As measured, the 64.4 vs 53.5 gap is a whole-system property; it cannot distinguish a contribution from the accurate DexJoCo targets (an auxiliary-task or regularization effect at fine-tuning) from a contribution of the estimated human-video targets, or even rule out that the estimated targets add noise the model must unlearn. The conclusion's inference that the world-modeling benefit is carried by the human-video scene forecasts is therefore untested. A 2x2 ablation (scene supervision on/off per stage), or removal of scene supervision from a single stage at a time, would resolve the attribution; without such an experiment, the paper's abstract-level claim that scene-trajectory supervision accounts for 10.9 points is overstated relative to what the evidence supports.","section":"§4.4, Table 3(b); §3.4; App. C.1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The '2D frame' comparison (60.1 vs 64.4) varies two things at once: the forecast target (DINOv3 image latents versus 3D point trajectories) and the encoder (the DINOv3 image encoder versus the Mosaic3D point encoder). Since Table 3(a) shows a 5.4-point gap between exactly these two encoders on the 3D-point pipeline (59.0 with lifted DINOv3 features versus 64.4 with Mosaic3D), part or all of the 4.3-point gap in panel (b) may be an encoder effect rather than an effect of the trajectory-target form. The sentence 'Forecasting the 3D trajectories of scene points from 3D features thus outperforms forecasting future image latents from image features' attributes the gap to the target form alone, which the reported cells do not isolate. A comparison that holds the encoder fixed (for example, forecasting the 2D latents with the Mosaic3D point encoder, or forecasting 3D point targets with the DINOv3/PointWorld-style lifted features) would support the stated attribution.","section":"§4.4, Table 3(b), 2D frame row","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The 56.9-point pre-training gain in the abstract uses the full-corpus endpoint (12.1% to 69.0%), while Fig. 6 reports 60.1% for EgoDex alone and Fig. 7 shows 69.0% only after adding VITRA. The numbers are consistent, but the abstract should state that the 56.9-point figure corresponds to the combined EgoDex+VITRA corpus to avoid a mismatch for the reader who first sees Fig. 6.","section":"Abstract; §4.3, Figs 6-7","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Checkpoint selection for the V olt-H multi-task model is done on the first evaluation seed's ten-task average (from the 20k/40k/60k fine-tuning snapshots). This is a mild selection-on-evaluation step; the reported three-seed standard deviations would be cleaner if selection used a held-out seed or a validation split. The paper is transparent about the procedure, so this is a protocol nuance rather than an error.","section":"App. C.2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The palm-frame transfer to the ten Allegro keypoints via the fixed robot offsets d^k is a central modeling assumption, and no quantitative validation of the retargeting path (for example, correspondence error between human hand keypoints and the induced robot keypoints, or retargeting success on held-out clips) is reported. The paper states the assumption clearly, but a small validation metric would strengthen the claim that the ten-keypoint representation is the right transfer medium.","section":"Eq. (8); App. B.1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The real-robot comparison uses a single training run and 24 trials per method per task with no significance testing, so the reported margins (e.g., 50.0 versus 16.7/25.0 on square-peg insertion) should be read with that uncertainty in mind; the paper does state the trial counts, which is good. Separately, on RoboDojo-Precision the scene head is frozen because scene trajectories cannot be reconstructed (App. B.2), so those results exercise the hand-path and action components but not the scene-supervision mechanism and should not be cited as evidence for the scene-supervision claim.","section":"Real-robot evaluation, Fig. 5; RoboDojo, App. B.2","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"What's new: PointWAM jointly forecasts scene points and hand keypoints as 3D point trajectories in one shared space-time frame and retargets the hands directly, no separate planner. That is genuinely new relative to EgoWAM (grippers, no fingers) and PointWorld/PointZero (scene only, no actions). Pre-training on 1.15M human-video episodes lifts DexJoCo success from 12.1% to 69.0%, beating GR00T N1.6 by 11.7 points. The ablation supporting that pre-training gain is clean: same pipeline, no-pre-training versus with, internal baselines measured identically, scaling curves with data and model size. The paper is honest about its own caveats and states the checkpoint-selection protocol explicitly.\n\nThe stress-test point holds, and it is the sharpest issue. The 10.9-point scene-supervision gain (Table 3b) is ablated by removing scene supervision from pre-training and fine-tuning simultaneously. The two stages use different targets: accurate simulator replay segmentation at fine-tuning, unvalidated CoTracker3/VGGT-Omega estimates at pre-training. So the gain cannot be attributed to the human-video scene trajectories; it could live entirely in the fine-tuning stage as an auxiliary-task effect. Validating the estimates would not settle it. The broader \"largely from human videos\" conclusion still has support from the clean pre-training ablation, but the scene-forecasting component of that transfer specifically is unpinned. One clean split — remove scene supervision from pre-training only, then from fine-tuning only — would resolve it.\n\nOther soft spots, in proportion. The 2D-versus-3D world-modeling row swaps in a different vision encoder (DINOv3), so \"3D beats 2D\" is confounded with encoder choice; the paper says this itself, but it still undercuts the claim as stated. Real-robot numbers are 24 trials per method, no error bars — a real result on a hard bimanual insertion, just thin. Checkpoint selection on the first seed's ten-task average is transparent and probably modest in effect. None of these touches the core pre-training claim.\n\nBottom line: the representation and the human-video transfer are real contributions, the SOTA numbers are plausibly supported, and the abstract's scene-supervision attribution needs the disaggregated ablation before the headline holds together. This deserves a serious referee, with a conditional path to accept.","headline":"Real advance in dexterous manipulation with a genuinely new joint scene-plus-hand 3D trajectory formulation and clean pre-training ablations, but the headline scene-supervision gain conflates accurate fine-tuning targets with unvalidated human-video estimates, and the 2D-vs-3D comparison is encoder-confounded.","tokens_in":26810,"tokens_out":2968,"duration_ms":19445},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"PointWAM claims that jointly forecasting the scene and hands as 3D point trajectories in a shared space-time frame, pre-trained on 1.15M human demonstration videos and fine-tuned on robot demos, raises DexJoCo success from 12.1% to 69.0%…","keywords":["world action model","dexterous manipulation","3D point trajectories","human video pre-training","scene forecasting","action retargeting","point cloud policy","world modeling"],"falsifier":"Recompute DexJoCo scene trajectories from the simulator's known rigid-body ground truth instead of the paper's CoTracker3-plus-VGGT-Ω pipeline, and re-run the scene-forecasting ablation of Table 3b: if the 10.9-point gain shrinks or vanishes with validated supervision, the world-modeling claim is partly an artifact of the estimator's bias. A second decisive check is to re-pretrain on human videos whose scene depth is established by multi-view reconstruction with independent metric scaling (stereo or known object sizes), and test whether the 56.9-point pre-training gain survives.","tokens_in":27044,"feed_emoji":"🤖","tokens_out":8797,"duration_ms":50873,"temperature":0.7,"pith_summary":"PointWAM claims that a robot can acquire dexterous manipulation largely from human demonstration videos, as long as the world is represented as 3D point trajectories — scene and hands in a single space-time frame — and the model is trained to forecast how those points will move before retargeting the forecast hands into actions. Pre-training on 1.15M human episodes from EgoDex and VITRA raises average DexJoCo success from 12.1% to 69.0%, a 56.9-point gain, and supervising the scene's motion adds another 10.9 points over forecasting the hands alone. With both, PointWAM beats the prior best multi-task success rate on DexJoCo by 11.7 points and outperforms robot-pre-trained VLAs on a real robot, including a task with a table-colored object that color-based baselines miss. If correct, the claim would show that explicit 3D scene forecasting from human video is a stronger recipe for dexterous control than image-latent world modeling or vision-language-action policies trained on robot data.","feed_headline":"Human videos add 57 points to dexterous robot success","feed_subtitle":"PointWAM forecasts scene and hands as 3D point tracks and retargets hand motion, beating the prior best by 11.7 points.","key_machinery":"The load-bearing object is the joint 3D point-trajectory forecast in a shared space-time frame. Scene points and hand keypoints are tokenized onto a common 1 cm voxel grid, aggregated into patches of $K_v^3$ voxels, and processed by a transformer encoder with axial 3D rotary position embeddings; the decoder then emits per-point displacement forecasts for scene and hands from the same features. A frozen Mosaic3D encoder supplies text-aligned 3D point features, and a separate transformer decoder — the action retargeter — turns predicted hand trajectories plus robot state into an action chunk of end-effector poses and finger joints. The mechanism does three jobs at once: it forces the backbone to learn how scene and hand points co-evolve, it lets human videos (which have no robot actions) supply full supervision, and it removes any task-specific object or keypoint selection, since every observed scene point is forecast.","core_discovery":"PointWAM claims that decomposing the world into a scene and hands, and forecasting both as 3D point trajectories in one shared space-time coordinate frame, is a representation that transfers dexterous skill from human video to robot control. Given a colored point cloud $P_t$ and a language instruction, the model jointly predicts the future displacement of every observed scene point and of ten hand keypoints per hand, $u^{(i)}_{t,\tau} = p^{(i)}_{t+\\tau} - p^{(i)}_t$ and $v^{(k)}_{t,\tau} = q^{(k)}_{t+\\tau} - q^{(k)}_t$, then a retargeter maps the forecast hand trajectories plus robot state into an action chunk. Because human and robot demonstrations collapse to the same trajectory format, one backbone is pre-trained on 1.15M human episodes and fine-tuned on robot demos. The reported numbers: the pre-training data raises DexJoCo success from 12.1% to 69.0%, scene-trajectory supervision adds 10.9 points over forecasting hands alone, and the full model beats the strongest baseline, GR00T N1.6, by 11.7 points and beats robot-pre-trained VLAs on a real robot, with the largest margin on an out-of-domain object whose color resembles the table.","pith_inferences":["My inference: if the scene-forecasting gain is real, it suggests contact geometry — not appearance — is what dexterous policies most need to anticipate; a natural extension the paper leaves implicit is feeding the forecast scene contact points into the retargeter, which could reduce grasp slippage in tight insertion tasks where the paper's failures cluster.","My inference: the strongest risk to the world-modeling claim is supervision quality — the scene trajectories in human video are estimates (CoTracker3 2D tracks lifted by VGGT-Ω monocular depth scaled to hand annotations), so the scene-forecasting gain could be partly learning the estimator's bias rather than true dynamics; this is testable, because DexJoCo's simulator provides known rigid-body geo","My inference: forecasting every scene point without keypoint selection suggests the approach could generalize to other manipulation modalities where contact points are not known in advance, such as legged manipulation or tool use, where no human annotator can pre-specify the relevant scene points.","My inference: the real-robot evaluation is small (24 trials per task), so the most decision-relevant number to reproduce independently is the out-of-domain banana result (58.3% versus 16.7% and 25.0%), because it claims geometry beats color when appearance cues are ambiguous."],"forward_implications":["Large-scale human video becomes a viable pre-training corpus for dexterous robotics, reducing reliance on robot demonstrations that are expensive to collect, since only a small robot fine-tuning set (959 DexJoCo demos) is needed after pre-training.","Forecasting the future scene in 3D points beats forecasting it as image latents for dexterous control: the 3D variant gains 10.9 points over hands-only forecasting, while the 2D-latent variant gains 6.6 and still trails the 3D model by 4.3 points, suggesting the gain is specific to 3D trajectory supervision, not world modeling per se.","The point-trajectory representation is not tied to hands: the forecaster transfers to parallel-jaw grippers on RoboDojo-Precision, where human-hand pre-training reaches 8.8 average score and 4.8% success against near-zero scores for video-pretrained Fast-WAM and PointACT, and beats robot-gripper-pretrained GR00T N1.6.","Performance scales with both human-demonstration volume (60.1% with EgoDex alone rising to 69.0% with the full 1.15M-episode corpus) and trainable backbone capacity (52.3% at 39M parameters rising to 69.0% at 655M), so further gains may follow from more human data and larger models.","Under full randomization of camera pose, lighting, and texture, the single-task margin over the strongest baseline widens from 5.3 to 22.2 points, indicating that forecasting 3D scene dynamics is more robust to visual variation than RGB-based methods.","The 3D encoder (Mosaic3D) outperforms lifted image features (DINOv3) by 5.4 points, indicating that encoding 3D structure in the features themselves, rather than only in their positions, matters for dexterous manipulation."],"supporting_citations":[{"why":"Supplies the EgoDex egocentric human videos with ARKit hand annotations that form the core of the 1.15M-episode pre-training corpus.","marker":"Hoque et al., 2026"},{"why":"Supplies the VITRA human videos with reconstructed hand poses that extend the pre-training corpus and add 8.9 points of success.","marker":"Li et al., 2026b"},{"why":"Provides the DexJoCo benchmark, the replay-based scene-trajectory reconstruction, the baselines (DP-T, pi0.5) it beats, and the evaluation protocol.","marker":"Wang et al., 2026a"},{"why":"Supplies CoTracker3 2D tracks used to extract scene trajectories from both human videos and real-robot recordings.","marker":"Karaev et al., 2025"},{"why":"Supplies VGGT-Ω monocular depth that lifts the 2D scene tracks into 3D, with metric scale recovered from hand annotations.","marker":"Wang et al., 2026b"},{"why":"Supplies the frozen Mosaic3D text-aligned 3D point encoder whose features outperform lifted image features by 5.4 points.","marker":"Lee et al., 2025"},{"why":"Supplies the Volume Transformer backbone and patch-tokenization scheme that the trajectory forecaster builds on.","marker":"Yilmaz et al., 2026"},{"why":"Supplies GR00T N1.6, the robot-pre-trained VLA baseline that PointWAM must beat by 11.7 points on DexJoCo.","marker":"NVIDIA et al., 2025"},{"why":"Supplies the scene-loss weighting scheme that PointWAM adopts for its scene-trajectory supervision, plus the bounds-filter cropping idea.","marker":"Huang et al., 2026"}],"fun_headline_variants":["PointWAM maps 3D point tracks to boost dexterous robot skill","Human video pre-training lifts DexJoCo success by 57 points","3D world action model forecasts scene and hands for robots","PointWAM beats prior best by 11.7 points with point trajectories","Forecasting 3D point motion transfers human manipulation to robots"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The gain credited to scene forecasting (10.9 points) rests on scene trajectories that are never validated: CoTracker3 2D tracks lifted to 3D with VGGT-Ω monocular depth, whose metric scale is recovered by assuming the ARKit and VITRA hand annotations are accurate, so systematic error in those estimates would shrink or erase the world-modeling contribution.","fun_headline_variants_meta":{"raw":{"variants":["PointWAM maps 3D point tracks to boost dexterous robot skill","Human video pre-training lifts DexJoCo success by 57 points","3D world action model forecasts scene and hands for robots","PointWAM beats prior best by 11.7 points with point trajectories","Forecasting 3D point motion transfers human manipulation to robots"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000165,"raw_usage":{"cost":0.000165368,"is_byok":false,"cost_details":{"upstream_inference_cost":0.000165368,"upstream_inference_prompt_cost":0.000123368,"upstream_inference_completions_cost":4.2e-05},"total_tokens":1236,"prompt_tokens":1086,"completion_tokens":150,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":56}},"tokens_in":1086,"tokens_out":150,"cost_source":"openrouter_usage","duration_ms":2190,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":56,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-10-05T17:22:35.803860+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Recompute DexJoCo scene trajectories from the simulator's known rigid-body ground truth instead of the paper's CoTracker3-plus-VGGT-Ω pipeline, and re-run the scene-forecasting ablation of Table 3b: if the 10.9-point gain shrinks or vanishes with validated supervision, the world-modeling claim is partly an artifact of the estimator's bias. A second decisive check is to re-pretrain on human videos whose scene depth is established by multi-view reconstruction with independent metric scaling (stereo or known object sizes), and test whether the 56.9-point pre-training gain survives.","supporting_citations":[{"cited_title":"Yoon, Mouli Sivapurapu, and Jian Zhang","cited_arxiv_id":null,"evidence_quote":"Supplies the EgoDex egocentric human videos with ARKit hand annotations that form the core of the 1.15M-episode pre-training corpus."},{"cited_title":"CoTracker3 : Simpler and better point tracking by pseudo-labelling real videos","cited_arxiv_id":null,"evidence_quote":"Supplies CoTracker3 2D tracks used to extract scene trajectories from both human videos and real-robot recordings."},{"cited_title":"Mosaic3D : Foundation dataset and model for open-vocabulary 3D segmentation","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen Mosaic3D text-aligned 3D point encoder whose features outperform lifted image features by 5.4 points."},{"cited_title":"Volume Transformer : Revisiting vanilla transformers for 3D scene understanding","cited_arxiv_id":null,"evidence_quote":"Supplies the Volume Transformer backbone and patch-tokenization scheme that the trajectory forecaster builds on."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}