Literature that changes the traversal plan

Checked 18 September 2026. This is a targeted primary-source reassessment, not an exhaustive survey or a guarantee of novelty. Full-text access below means the relevant method, evaluation and limitation sections were inspected; it does not mean an external method was reproduced. Paper claims are attributed to their authors. Our design decisions are inferences.

The closest work now

Primary source and inspected scope What is established in that source Decision for KimoNav
TANGO, 2609.09158v1, §§3–5 and deployment/metrics appendices; full text RGB, language and proprioception condition whole-body reference chunks executed by SONIC. Its Plan–Edit–Track pipeline filters references through physical execution; action labels remain the edited references, not the tracked states. It also compares another tracker and identifies RGB-only perception and tracker capability as limitations. Direct overlap with our broad destination. Compare actual reference inputs and physical task definitions. “Whole-body VLA,” frozen-tracker composition, and swapping a tracker are insufficient novelty. Its outputs are references for a tracker, not direct motor torques.
PASSAGE, 2609.18732v1, §§III–V and map/runtime appendices; full text Destination and multi-layer geometric observations condition a flow planner and perceptive tracker. Scene-aligned data, chunk continuity and planner-side post-training support onboard traversal; goal completion and contact-free completion are distinguished. Essential new comparator. Scene-conditioned planning, layered overhead geometry, frozen-tracker refinement and reusable interfaces are already occupied. Evaluate our actual remaining question: task preservation and reliable adaptation under controlled command/observation changes.
CAT / HumanoidPF, 2601.16035v1, §§III–IV; full text Per-body geometric fields guide traversal learning; hybrid scenes cover ground, lateral and overhead obstacles. Deployment uses LiDAR–inertial mapping and target selection. Strong geometry/control comparator. Match sensing and motor capabilities before attributing a result to our interface. Per-body features and 3D clearance are established tools. Official repository was inspected for availability, not executed.
DWMP, 2609.12347v1, method and experimental sections; full text Separate dynamics and depth world-model representations feed a traversal policy, with simulation and G1 demonstrations. A predictive visual representation is a plausible later comparison, not a necessary first component. Compare with a simple geometry/history encoder before adopting two world models.

These sources make a timing-only differentiator unconvincing unless a real application requires it. We should study an operationally meaningful failure, with matched baselines, rather than create a narrower score merely because existing systems were not evaluated on it. The proposed task/observation intervention study remains a hypothesis; this review does not establish that no prior work studies the same issue.

Motion, scene data, and contact

Source and scope Relevant finding Design implication
Moving Through Clutter, 2603.05993v1, §§III–V; full text VR collection couples geometry with whole-body traversal. The paper reports 348 trajectories across 145 scenes and discusses scene-agnostic retargeting and omitted contact-assisted progression. Connect our program to this existing scene-aligned resource. Preserve human motion, retargeted reference and executed robot data as separate levels. Reuse accessible assets before building another scene generator.
SceneBot, 2606.27581v1, §§3–6; full text Motion-conditioned scene reconstruction and contact labels train a scene-interaction tracker. Its conclusion leaves autonomous perceptive goal-driven control to future work. Strong contact-aware motor and data precedent. A task-level planner remains a different obligation. New support contacts require motor training or qualification, not only collision-free animation.
Perceptive BFM, 2606.08059v1, synthesis, distillation and limitations; full text Terrain-conformal reference synthesis and perceptual distillation adapt supplied motion. Its stated foot-centric assumptions can leave upper-body collisions. Use as a terrain-adaptation comparison. Preserving upper-body style is unsuitable when shoulder/head clearance requires changing that style.
Learning from Hallucination family, author project synthesis Open-space navigation experience is paired with constructed obstacle configurations; later variants change hallucination and deployment requirements. Motion-first scene construction is established. Our data branch must test whole-body feasibility, alternative actions and downstream learning against scene-first or existing scene-aligned data.
SUMMON, INFERACT, SceMoS, MotionBricks Previously reviewed in the September 15 hindsight/tokenization notes; not independently re-read in this revision. Keep as secondary prior art for scene reconstruction, motion vocabularies and root/body factorization. Verify the relevant full-text passages before a specific novelty claim.

What the behavior-model papers do and do not imply

Source and inspected scope Relevant finding Design implication
Behavior Foundation Model, 2509.13780v1, §§III–V; full text A tracking proxy, masked geometric control, CVAE and online student-state supervision provide reusable control. These are methodological precedents. Our offline context masking and limited stored queries are not a full reproduction. Judge the public prior independently of privileged reconstruction.
Scaling BFM, 2607.15163v1, architecture, scaling and deployment sections; full text plus existing source audit Time-offset motion targets and observation/action histories support global whole-body control; motion diversity and rollout quantity are distinct. Time tokens and cross-attention are established. Repeating the same clips cannot replace missing body/scene behavior. Relative features still need an explicit localization contract.
BFM-Zero, 2511.04131v1, formulation and inference; full text plus September 15 code audit Forward–backward learning supports objective-conditioned behavior and inference through a learned representation. A separate pretraining route. Its latent is not SONIC code64, and importing its name does not give our student zero-shot properties. Reconsider only against a specific task and compute budget.
SONIC, 2511.07820v4, control/planning and deployment sections; full text A motion tracking system and kinematic planning interface provide reusable humanoid motion execution. Keep both original and repaired checkpoints as explicit baselines. Reproduce the actual reference context, initialization and controller implementation; adapted pose replay does not establish full native deployment parity.
ADAPT, 2609.00677v1, §§3–4; full text Closed-loop text-driven control, robustness adaptation and goal steering include a reaching-and-stationary objective. Neither text control nor stopping is new. Compare command semantics, allowed observations and execution latency. Our strict historical timing results are not directly comparable with its goal criterion.
MaskedMimic, HOVER, CLAW, Humanoid-LLA, GenTrack, RTC Covered in existing September 8/15 primary-source reviews; those local reviews were inspected here. No new full-text verification in this revision. Retain as comparator families for sparse control, parameterized planning, language and execution-aware generation. Check implementation and availability before selecting an experimental baseline.

A focused next reading pass

  1. Task and runtime contracts: inspect released TANGO/PASSAGE/CAT code, if available, for goal acceptance, body/reference representation, collision accounting, recovery, sensor delay and initialization. Do not treat a project page as proof that a runnable checkpoint exists. This revision does not certify the release completeness of TANGO or PASSAGE.
  2. Task preservation under partial observation: seek body-aware planning and constrained/uncertainty-aware navigation methods with explicit false-admission, abstention and intervention outcomes. Include conventional search/MPC and geometric checks, not only learned policies.
  3. Supervision that distinguishes decisions: compare scene-aligned demonstration, scene-first generation, and motion-first augmentation at matched real behavior coverage. Ask whether changing an obstacle changes the desired action under a fixed destination.
  4. Compression only after capability: revisit online masked distillation and multimodal prediction after a planner–tracker provides useful actions. Test whether compression improves latency/data cost while retaining task and contact outcomes.

Stop a reading branch when it resolves the design decision. Continue the nearest-work search before a paper claim, especially because the September 2026 traversal literature has changed the comparison set substantially.