FootQuery: Future-Touchdown-Guided Retrieval from Depth History for Perceptive Humanoid Locomotion
Tao Dong, Jia Yu, Yuxuan Fan, Linna Zhao, Jiaqi Gong, Andong Yang, Chao Gao, Guyue Zhou
cs.RO · cs.LG · eess.SY
Quick verdict
FootQuery predicts each foot's touchdown to query depth history for humanoid locomotion; simulation evidence is coherent but statistically thin, hardware results qualitative.
Summary
FootQuery is a perceptive humanoid locomotion framework that queries stored depth history using each foot's predicted next touchdown. Eight historical depth frames are encoded into 288 spatial tokens plus a recurrent global memory; a proprioceptive head predicts per-foot touchdown location and uncertainty, which condition cross-attention queries over the tokens. During training, realized contacts are projected into historical frames to supervise retrieval. Training is further supported by a progressive force-assistance curriculum and tread-midline shaping. The complete framework beats its ablations at the largest obstacles (87% on 20-cm stairs, 94% on 50-cm gaps, 91% on 50-cm platforms), though HPL matches it on the 50-cm gap; against released MoRE and Hiking controllers it attains 100% success on all six terrains with only 10 trials per terrain. A diagnostic on stair traversal reports 7.82% of future touchdowns visible in the current ROI versus 67.89% in some retained frame, with attention concentrating on historically visible regions; XY prediction error grows from 1.31 cm at a 3-step lead to 8.52 cm at 24 steps. A single policy runs on a Unitree G1 across outdoor stairs and indoor mixed routes. Statistical reporting is sparse (no seeds, CIs, or trial counts for ablations) and hardware results are qualitative.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Predicting each foot's next touchdown from proprioception and using that distribution to query depth history gives the policy access to terrain evidence that has left the current field of view (motivated by 7.82% current-ROI vs 67.89% any-frame visibility in Sec. IV-D).
- The complete framework outperforms its component ablations on the most challenging tested settings (87%/94%/91% on 20-cm stairs / 50-cm gaps / 50-cm platforms, Fig. 7).
- Progressive force assistance accelerates early optimization by roughly 1.5-2 thousand PPO iterations to 0.5 aggregate success, with the gap narrowing late in training (Fig. 8).
- A single policy, using only proprioception and onboard depth, traverses outdoor stairs and indoor mixed stair/platform/gap routes on a Unitree G1 (Sec. IV-F).
Strengths
Novel, well-motivated retrieval mechanism
Sec. I/III condition depth-history queries on each foot's predicted touchdown distribution rather than current state or reference motion (contrast CReF [16], DELTA [17]); the left/right feet query separately, matching their different landing sites and timings.
Honest, quantitative diagnostic section
Sec. IV-D quantifies the core premise: current-ROI visibility 7.82% vs 67.89% in any retained frame (60.07% history-only), attention positive-set mass 34.30% vs 5.58% uniform, GT-best-head Top-1/Top-3 81.73%/95.80% — the paper verifies its own motivation rather than asserting it.
Complete training recipe with unusual specificity
Clear acknowledgment of unresolved causality
Single-policy hardware demonstration across mixed terrains
Weaknesses
Underpowered ablation statistics
Baseline comparison overstates dominance
Diagnostic policy differs from deployed policy
Hardware evidence is qualitative only
Label-availability attrition is large
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
FootQuery (arXiv:2609.21447) proposes contact-conditioned retrieval over depth history for perceptive humanoid locomotion on the Unitree G1. The central idea: each foot's *predicted* next touchdown — a Gaussian (mu, sigma) over the pelvis-yaw XY plane predicted from proprioception alone (Eq. 2) — becomes a query over 288 tokens from eight historical depth frames (~1.4 s span), with attention supervised by projecting realized contacts into the historical frames where they were visible (Sec. III.B.4, Eq. 5). Two auxiliary training mechanisms — progressive force assistance (Sec. III.C) and tread-midline shaping (Sec. III.D) — are cleanly separated from the architecture. Claims are scoped honestly: superiority is asserted only "on the most challenging tested stairs, gaps, and platforms," and HPL is acknowledged to match the 94% gap result.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What problem does FootQuery solve in humanoid locomotion?
When a humanoid climbs stairs or crosses gaps, the spot where its foot will land has often already moved out of the camera's view or is hidden by its own legs by the time of contact. FootQuery stores recent depth images and retrieves the relevant old frame regions for each foot separately, using the predicted location of the foot's next touchdown as the query, so the policy can still exploit terrain it saw up to about 1.4 seconds earlier.
How does FootQuery predict where a foot will land?
A neural head takes current and historical proprioception (joint states, body orientation, previous actions) plus a foot-identity embedding and outputs a Gaussian distribution — mean and standard deviation — over the landing position of each foot in the pelvis-forward XY plane, bounded to about 1.0 m forward / 0.6 m lateral. During training the predictions are supervised with realized contact locations; at deployment the predicted distribution is embedded and used to form attention queries over historical depth-image tokens.
What results does the paper report?
In simulation the full method reaches 87% success on 20-cm stairs, 94% on 50-cm gaps, and 91% on 50-cm platforms, beating its own ablations; against released MoRE and Hiking controllers it scores 100% on all six tested terrains with 10 trials each. A diagnostic shows only 7.82% of future touchdowns are visible in the current camera ROI versus 67.89% in some stored frame, and touchdown-prediction error grows from 1.31 cm at a 3-step lead to 8.52 cm at 24 steps. The same policy was demonstrated on a real Unitree G1 on outdoor stairs and indoor mixed-terrain routes, though these hardware results are qualitative.
Is code or a trained model available for FootQuery?
The paper does not mention a code or model release, so reproduction currently requires reimplementation. It does provide an unusually detailed training recipe (MuJoCo/mjlab, single RTX 5090, exact loss weights, force-assist schedule, reward shaping parameters), which lowers but does not eliminate the reimplementation barrier.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 7-11 are only available as summarized text; curve shapes, per-difficulty ablation points, and qualitative contact plots could not be independently inspected.
- Tables I and II were fully extracted and anchor the quantitative claims, but no error bars exist in the source text itself.
- Hardware experiments (Sec. IV-F) are described qualitatively; no trial counts or success statistics exist to assess.
- Several related works (e.g., SOLO, Hiking, SSR) are 2025-2026 arXiv preprints that could not be verified for this review.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.