{"id":"4a9cc669-b723-4460-937d-9770dfbcb68b","arxiv_id":"2605.20811","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Demo-JEPA enables one-shot cross-embodiment imitation by mapping visual demonstrations to shared latent future trajectories that serve as subgoals for the target agent's own forward dynamics planning.","lead":"Demo-JEPA uses a joint-embedding predictive architecture to turn visual demonstrations from one robot or human into latent future trajectories that a different target robot can plan toward using its own dynamics model. This could let robots imitate tasks across mismatched bodies and action spaces without retargeting or shared action data.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"JEPA encoder trained only on target visuals may not embed source demonstration frames into a latent space whose predicted trajectories are realizable under target dynamics.","rationale":"The reader's weakest assumption correctly isolates the cross-embodiment transfer step. The concrete test above directly probes whether the encoder produces compatible latents without any additional alignment mechanism, which is the precise point where the central claim could fail even if the target planner itself works.","tokens_in":1727,"tokens_out":347,"duration_ms":17993,"concrete_test":"Extract the visual encoder from the trained JEPA, run it on held-out source demonstration frames and on matched target frames of the same semantic state, then compare the cosine distance between the resulting latents; if source latents lie outside the convex hull of target latents by more than the intra-target variance observed during training, the shared-space assumption is violated.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The framework trains the JEPA world model exclusively on the target agent's own interaction data, so the visual encoder and predictor are optimized for target-specific appearance and dynamics. Source demonstrations are then encoded with this same encoder and rolled out in the shared latent space to produce subgoals. For the translation step to succeed, the encoder must map visually dissimilar source frames (different morphology, gripper, camera angle) to latents whose subsequent predictions under the target predictor correspond to states the target can actually reach. No explicit cross-embodiment alignment, contrastive loss, or domain-adaptation term is described that would enforce this invariance; success therefore rests on the untested assumption that the target's learned visual features are already sufficiently semantic and embodiment-agnostic.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript introduces Demo-JEPA, a cross-embodiment imitation framework built on a JEPA world model. Source visual demonstrations are encoded into a shared predictive latent space to produce future trajectories that serve as subgoals; the target agent then plans to realize these trajectories using its own learned forward dynamics. The approach requires only visual demonstrations and the target agent's interaction data, avoiding action retargeting or multi-embodiment co-training. Experiments on RLBench and real-world manipulation are reported to match in-domain planners while generalizing to unseen tasks and embodiment variations.","tokens_in":1862,"tokens_out":530,"duration_ms":24030,"significance":"If the central mechanism holds, the work offers a principled way to separate intent inference from embodiment-specific execution in imitation learning. The reliance on a predictive JEPA representation rather than direct action matching is a conceptual strength, and the one-shot setting with only target self-interaction data could reduce data requirements compared with prior cross-embodiment methods. Reproducible code or explicit falsifiable predictions are not mentioned in the provided text.","major_comments":[{"comment":"Method section (description of JEPA training and inference): the claim that a JEPA encoder and predictor trained exclusively on target-agent interaction data produces a latent space in which source demonstration frames yield realizable target trajectories rests on an untested invariance assumption. No contrastive loss, domain-adaptation term, or explicit cross-embodiment alignment is described; therefore the translation step may map visually dissimilar source frames to latents whose predicted futures are unreachable under the target dynamics. This assumption is load-bearing for the central cross-embodiment claim.","section":"Method"},{"comment":"Experiments section (quantitative results and ablations): the abstract asserts that Demo-JEPA matches specialized in-domain planners and generalizes where prior methods fail, yet no success rates, baseline comparisons, ablation studies on the predictive component, or metrics for latent-space alignment across embodiments are supplied in the available text. Without these, the empirical support for generalization cannot be evaluated.","section":"Experiments"}],"minor_comments":[{"comment":"Notation for the latent trajectory and subgoal extraction should be defined explicitly with equations rather than prose only.","section":"Method"},{"comment":"Figure captions for any qualitative rollout visualizations should include embodiment labels and camera viewpoints to clarify cross-embodiment differences.","section":"Experiments"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for their constructive review and for identifying key areas where the manuscript could be strengthened. We address each major comment below with clarifications on the method and plans to improve the experimental presentation.","responses":[{"response":"We thank the referee for this observation. The JEPA world model is trained exclusively on the target agent's self-supervised interaction data using a predictive objective that learns to forecast future latent states. This training encourages the encoder to produce representations focused on task-relevant dynamics rather than embodiment-specific visual features, as the loss penalizes inaccurate future predictions under the target's own actions. Consequently, when source demonstration frames are encoded into this space, the resulting latent trajectories correspond to future states that the target can realize by planning with its learned dynamics model. No explicit alignment term is included by design, since the approach avoids requiring paired cross-embodiment data. We will revise the method section to expand on this rationale, including why the predictive (rather than reconstructive) objective supports the observed cross-embodiment generalization.","revision_made":"partial","referee_comment":"[Method] Method section (description of JEPA training and inference): the claim that a JEPA encoder and predictor trained exclusively on target-agent interaction data produces a latent space in which source demonstration frames yield realizable target trajectories rests on an untested invariance assumption. No contrastive loss, domain-adaptation term, or explicit cross-embodiment alignment is described; therefore the translation step may map visually dissimilar source frames to latents whose predicted futures are unreachable under the target dynamics. This assumption is load-bearing for the central cross-embodiment claim."},{"response":"We agree that the quantitative support should be presented more explicitly. While the manuscript reports results on RLBench and real-world tasks showing performance comparable to in-domain methods and superior generalization, we will revise the experiments section to include a dedicated table of success rates, direct numerical comparisons to baselines (such as behavior cloning and other cross-embodiment approaches), an ablation isolating the contribution of the predictive JEPA component, and metrics evaluating latent trajectory consistency across embodiments. These additions will make the empirical claims fully evaluable.","revision_made":"yes","referee_comment":"[Experiments] Experiments section (quantitative results and ablations): the abstract asserts that Demo-JEPA matches specialized in-domain planners and generalizes where prior methods fail, yet no success rates, baseline comparisons, ablation studies on the predictive component, or metrics for latent-space alignment across embodiments are supplied in the available text. Without these, the empirical support for generalization cannot be evaluated."}],"tokens_in":1441,"tokens_out":554,"duration_ms":56843,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"Demo-JEPA reframes cross-embodiment imitation around JEPA latent trajectories as subgoals instead of action mapping, which is a clean idea but rests on an untested assumption that target-only training produces usable shared representations. The paper shifts the problem from reproducing embodiment-specific actions to inferring future states the demonstrator is aiming for. They train a JEPA world model on the target agent's interaction data, encode source visual demonstrations into the same latent space, roll out predicted trajectories as subgoals, and let the target plan to them with its own dynamics model. This setup needs only visuals from the demo plus the target's experience, avoiding retargeting or shared action spaces. That is a practical advantage when hardware varies. The abstract reports that experiments on RLBench and real manipulation tasks show it matches in-domain planners and generalizes to unseen tasks and embodiment changes where prior methods fail. The framing earns credit for treating demonstrations as goal specifications rather than action sequences and for extending JEPA-style predictive representations to one-shot transfer. The results, as described, suggest the approach can handle heterogeneous embodiments without large-scale co-training. The main soft spot is the assumption that the JEPA encoder, fit only to target visuals and dynamics, will map dissimilar source frames to latents whose predictions remain reachable under the target model. No explicit alignment, contrastive term, or domain adaptation is described, so the method depends on the learned features already being semantic and embodiment-invariant enough. The stress-test note correctly flags this risk; if the full experiments lack targeted ablations on cross-embodiment embedding quality or failure modes, the generalization claims would need more support. This work is aimed at robotics researchers focused on imitation learning and world models. A reader dealing with transfer across different robots or human-to-robot settings would find the problem statement and the subgoal formulation useful to consider. It deserves a serious referee because the core idea is distinct from existing action-centric methods and the problem is relevant, even though the evidence for the shared latent space would likely be the main point of scrutiny in review.","headline":"Demo-JEPA reframes cross-embodiment imitation around JEPA latent trajectories as subgoals instead of action mapping, which is a clean idea but rests on an untested assumption that target-only training produces usable shared representations.","tokens_in":2342,"tokens_out":500,"would_cite":false,"duration_ms":40811,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":{"model":"grok-4.3","evidence":[],"headline":"Robotics imitation via JEPA latent planning has no structural overlap with RS forcing chain","alignment":"orthogonal","rationale":"Paper centers on cross-embodiment goal inference in a JEPA world model with cross-attention fusion and CEM planning; no reference to J-cost, ratio symmetry, φ-ladder, 8-tick periodicity, or distinction-forced constants. Domain (cs.RO) lies outside RS theorems on recognition cost and spacetime emergence.","tokens_in":54898,"confidence":"high","tokens_out":116,"duration_ms":10889,"cache_read_input_tokens":38528,"cache_creation_input_tokens":0},"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A JEPA world model turns visual demonstrations into latent subgoals that any robot body can plan toward.","keywords":["cross-embodiment imitation","joint-embedding predictive architecture","imitation learning","world models","latent trajectories","robotic manipulation","one-shot learning","goal inference"],"falsifier":"Run a controlled test where the visual demonstration shows one clear goal but the translated latent trajectory leads the target robot to a different outcome; consistent failure to match the demonstrator's intent would falsify the shared-space claim.","tokens_in":2627,"feed_emoji":"🤖","tokens_out":623,"duration_ms":42511,"temperature":0.7,"pith_summary":"The paper reframes imitation as inferring intended future states from visuals rather than copying actions that depend on body shape or controls. A JEPA-based world model trained on the target robot's own experience creates a shared space where source demonstrations become future latent trajectories. The target robot treats those trajectories as subgoals and reaches them by planning with its own forward model. This matters because it removes requirements for matching action spaces or training across many robot types at once, needing only one visual demo plus the target's self-collected data. Experiments on RLBench and real manipulation tasks show the approach matches specialized planners while succeeding on unseen tasks and body changes where earlier methods break.","feed_headline":"Latent trajectories turn visual demos into robot subgoals","feed_subtitle":"A shared predictive space maps source intent to target plans without action matching or multi-body training.","key_machinery":"The JEPA-based world model that builds a shared predictive representation space to convert visual demonstrations into future latent trajectories usable as subgoals by the target agent.","core_discovery":"Demo-JEPA translates source visual demonstrations into target-compatible future latent trajectories in a shared predictive representation space. The target agent then uses these latent trajectories as subgoals and realizes them through planning under its own learned forward dynamics. Because Demo-JEPA avoids action-level correspondence and requires only visual demonstrations plus the target agent's own interaction experience, it supports flexible imitation across heterogeneous embodiments.","pith_inferences":["The same latent space might let a single world model translate goals between agents that differ even more, such as from human hands to robotic grippers.","Predictive models could become standard translators for goal inference in multi-robot teams with mismatched sensors.","Further tests with extreme morphology gaps would show where the shared representation starts to lose intent information."],"forward_implications":["Imitation succeeds without shared action spaces, retargeting, or multi-embodiment co-training.","The method generalizes to unseen tasks and new embodiment configurations.","Performance matches specialized in-domain planners on RLBench and real-world manipulation tasks.","Only visual demonstrations and the target agent's own experience are required."],"fun_headline_variants":["Demo-JEPA maps visual demos to latent subgoals","Predictive latents bridge different robot embodiments","Visual demos become subgoals without action alignment","JEPA world model supports flexible cross-embodiment learning"],"cache_read_input_tokens":64,"weakest_assumption_plain":"A world model trained primarily on the target agent's interactions can still produce a latent space that correctly captures the intent behind demonstrations from other embodiments.","fun_headline_variants_meta":{"raw":{"variants":["Demo-JEPA maps visual demos to latent subgoals","Predictive latents bridge different robot embodiments","Visual demos become subgoals without action alignment","JEPA world model supports flexible cross-embodiment learning"]},"model":"grok-4.3","cost_usd":0.007781,"raw_usage":{"total_tokens":3547,"prompt_tokens":654,"num_sources_used":0,"completion_tokens":59,"cost_in_usd_ticks":77812000,"prompt_tokens_details":{"text_tokens":654,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2834,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":654,"tokens_out":59,"duration_ms":37913,"temperature":1.0,"reasoning_tokens":2834,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-05-21T04:41:52.015014+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Run a controlled test where the visual demonstration shows one clear goal but the translated latent trajectory leads the target robot to a different outcome; consistent failure to match the demonstrator's intent would falsify the shared-space claim.","supporting_citations":[],"review_version":1}