ONGOING RESEARCH · SEPTEMBER 2026
KimoNav studies how a small, path-conditioned adapter can learn from generated humanoid motion. We keep the motion model frozen and ask a harder question: does the adapter use the requested path, or just learn a generic correction?
Exploratory offline development results. The navigation objective is longer-term; these experiments measure motion reconstruction, with privileged achieved root trajectories. They do not establish autonomous navigation or physical feasibility.
Walk, turn, walk again. Vary the approach distance, speed, and signed angle while keeping the generation procedure fixed. The resulting clips supply motion targets and expose failures worth learning from.
The text prompt describes steady walking. A structured program supplies the requested root path: three approach distances, two speeds, and six signed turn angles (±60°, ±90°, ±120°). Two generation seeds produce 72 clips. Turn-aware windows put the heading change inside the evaluated interval.
All 72 clips have the requested turn sign, but 50 fail the original terminal-speed check. A one-second hold is included; it does not erase those failures. The benchmark retains every program and diagnostic flag (bank audit). The 36 programs share six first-command prefixes, so they are not 36 independent environments.
Loading the recorded experimental results…
Joint reconstruction error reduction vs. native ARDY
Positive is better. Aggregate joint MSE reduction (%), not navigation success. Equal-weight training seeds; source windows averaged before programs.
| Method / control | MSE reduction | Joint RMSE | Contact accuracy |
|---|
Joint RMSE is the square root of aggregated coordinate MSE, shown in mm. Contact accuracy compares decoded contact labels to the source motion; it is not physical contact validation. Noise levels are fixed teacher timesteps, not elapsed motion time. The per-example oracle is privileged optimization headroom, not a learned deployable model.
All thresholds were fixed before this run. Passing a reconstruction gate would still require separate adapted-generation and navigation evaluations. Read the frozen execution contract ↗
Metric commands compile into requested position, heading, speed, and goal features.
Interpolate eight path nodes to thirteen generated token centers; preserve angle continuity.
A 78,592-parameter head changes temporal body features. ARDY parameters stay frozen.
Compare with time-only, TRAIN-mean, shuffled-path, and phase-shifted controls.
The earlier head sees the same compressed whole-path vector at every token. The new representation gives each token the path features for its own interval. The head size, 64 active path channels, initialization, draw sequences, and update budget remain matched. Both temporal alignment and geometric scope change, so this experiment cannot isolate their individual effects.
The raw loss regressed at low noise. A fixed, TRAIN-calibrated weighting repaired that aggregate, but a tiny contact-accuracy decrease kept its repair gate closed. We test local features under both the raw and balanced objectives, without selecting a favorable branch after seeing the outcomes.
Build the 72-clip turn-event bank. Per-example oracle fitting reveals 72.85% turn-joint MSE headroom; this measures fitting potential, not generalization.
Turn-joint MSE improves 1.65% over native ARDY. The matched time-only and TRAIN-mean controls prevent a useful-path claim.
Turn gain at t = 0 changes from −24.00% to +0.61%. Overall turn gain reaches 3.78%; contact accuracy drops 0.011 percentage points vs. the raw head, and the TRAIN-mean control remains stronger.
Nine shared regression prefixes, eighteen matched decoded refinements, and four score modes per objective. Full results above.
The next method is chosen from the complete matched comparison, with the failed controls retained.
The latest experiment uses three held-out approach-distance folds and training seeds 17, 29, and 43. Each of 1,020 new teacher examples is held out once per mode and seed. There are 24,480 new score rows and 30,600 cached comparison rows, plus 2,040 cached native/oracle references. This is a development benchmark; the original development partition is excluded and the original test partition remains unused.
Each shared regression prefix receives 1,000 updates, then branches into two 500-update refinements. Normalizers and templates use fold TRAIN data only. An independent validator reconstructs them and checks saved splits, draws, update counts, and cached aggregates. A 500-update replay reproduced the prior parameters and 312 score rows exactly. All 18 zero/disabled recovery checks are implementation tests, not evidence of better adapted generation.
The model parameters and recorded diffusion buffers remain unchanged. Historical native references used fresh CPU-computed schedule buffers; a separate audit found negligible aggregate score differences after the CUDA refresh, but replay showed that the distinction matters during fitting. This provenance is preserved in the report.
These experiments reconstruct held-out generated motion windows with privileged achieved root paths and fixed text. They do not measure scene collision avoidance, navigation success, long-horizon adapted generation, physical stability, hardware transfer, or open-vocabulary instruction understanding. Three training seeds and factorial programs do not establish statistical significance or environmental diversity.
KimoNav is ongoing work by Linji Wang. It builds on the pretrained motion models and implementations of ARDY and Kimodo from NVIDIA and their collaborators. The motion renders on this page use ARDY. These development results are not a published paper.