MamMA: A Mamba-Based Pedestrian Trajectory Prediction Algorithm Considering Occupancy Map and Pedestrian Awareness States
Juncen Long, Xiaofeng Jin, Gianluca Bardaro, Simone Mentasti, Matteo Matteucci
cs.CV · cs.LG
Quick verdict
MamMA: Mamba pedestrian prediction with LiDAR maps and awareness states; reported SOTA gains are large, but baselines are narrow and awareness labels assumed.
Summary
The paper introduces MamMA, a pedestrian trajectory prediction network that fuses three information sources: historical trajectories with pairwise pedestrian interactions, occupancy grid maps built from LiDAR (divided into patches and scanned by a Mamba model to extract obstacle features per pedestrian), and a discretized awareness state (normal vs. distracted) embedded alongside position features. Mamba state-space models replace graph convolutions along the temporal and spatial dimensions of a spatio-temporal graph, with separate map-scanning and decoding Mambas. On STCrowd, SiT, and JRDB (evaluated on validation sets, with awareness labels annotated by the authors), MamMA reports minADE3/minFDE3 of 0.27/0.49, 0.36/0.68, and 0.55/1.05 respectively — versus 0.38/0.69 for the best baseline on STCrowd — plus the lowest collision rates (0.87% vs at least 1.72% on STCrowd). Ablations (Tables III–V) confirm that maps and awareness states help, but also show the plain Mamba backbone ('Mam') alone already beats all compared baselines on STCrowd and on the reported ETH/UCY scenes. Main limitations: the seven baselines (2020–2025) all belong to the spatio-temporal graph-convolution family; no code release is stated; the awareness annotation protocol and any extractor from egocentric images are undescribed, so awareness is effectively a ground-truth input baselines lack; results are single-run without variance or significance tests; Table V omits the ETH scene despite the abstract's ETH/UCY claim; and on ETH/UCY, baselines use author-provided weights while MamMA is retrained.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 3
● low 2
4 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
- Scanning LiDAR-built occupancy maps patch-wise with a Mamba model extracts environment features that improve pedestrian trajectory prediction over methods relying on top-down imagery (Sec. I, III-D).
- Encoding a discretized pedestrian awareness state (normal vs. distracted) improves prediction, especially for distracted pedestrians and obstacle-adjacent samples (Sec. III-C; Table I Sdis/Sobs columns).
- MamMA achieves the best minADE3/minFDE3 and the lowest collision rate among seven compared methods on STCrowd, SiT, and JRDB (Table I).
- The Mamba backbone alone ('Mam') outperforms all compared baselines on STCrowd and on the reported ETH/UCY scenes even without maps or awareness inputs (Tables III, V).
- MamMA's accuracy degrades least near obstacles: minADE3 rises ~0.04 m (13.5%) versus ~0.08 m (18.3%) for the least-affected baseline as ground-truth-to-obstacle distance shrinks from 3 m to 0.5 m (Fig. 4).
Strengths
Sensor-realistic problem framing
Motivated by the observation (Fig. 1) that robots collect LiDAR plus egocentric images rather than top-down views, the paper builds occupancy grids from LiDAR (0.1 m resolution, Sec. IV-A) and adds a collision-rate metric (pcoll) alongside ADE/FDE, aligning evaluation with navigation safety rather than accuracy alone.
Robustness stratification is informative
The All/Sdis/Sobs splits (Sec. IV-A) and the distance-conditioned decay analysis (Fig. 4: as ground-truth-to-obstacle distance shrinks from 3 m to 0.5 m, MamMA's minADE3 rises ~0.04 m / 13.5% vs SGCN's ~0.08 m / 18.3%) directly probe the claimed obstacle and awareness benefits instead of relying on aggregate metrics.
Unusually complete architectural specification
Transparent ablations reveal the true driver
Weaknesses
Narrow baseline family behind the 'SOTA' claim
Awareness input is an unexplained oracle
Evaluation asymmetries
No statistical rigor
Headline novelty partially undercut by own ablations
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
MamMA (arXiv 2609.08041, cs.CV; Long, Jin, Bardaro, Mentasti, Matteucci — Politecnico di Milano) is a method paper on pedestrian trajectory prediction for mobile robots in human-robot coexistence environments. It claims three contributions (Sec. I): (1) a Mamba-based 'map scanning' module that extracts obstacle features from LiDAR-built occupancy grid maps patch-by-patch; (2) an embedding that injects a discretized pedestrian awareness state (normal vs. distracted) into position and interaction features; (3) a Mamba backbone replacing graph convolutions/attention along the temporal and spatial dimensions of a spatio-temporal graph. The headline claim is state-of-the-art minADE3/minFDE3 on STCrowd, SiT, JRDB, ETH and UCY, plus markedly lower collision rate (pcoll). The motivation (Fig. 1) is sound: robots collect LiDAR plus egocentric images, not top-down views, so environment features should come from occupancy maps rather than segmented overhead imagery.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 6 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 is MamMA?
MamMA is a pedestrian trajectory prediction model (Long et al., Politecnico di Milano, arXiv 2609.08041) that combines a Mamba state-space backbone with two extra inputs: occupancy grid maps built from LiDAR (scanned patch-by-patch for obstacle features) and a discrete pedestrian awareness state (normal vs. distracted). It predicts 4.8 seconds of future motion from 3.2 seconds of history.
How much better is MamMA than previous methods?
The paper reports minADE3 of 0.27 m on STCrowd versus 0.38 m for the best of seven compared baselines (~29% lower), with the lowest predicted-trajectory collision rate (0.87% vs at least 1.72%). The margins are large, but all baselines belong to the spatio-temporal graph-convolution family, results are single-run, and MamMA receives two inputs (map, awareness labels) the baselines do not use.
Which datasets does the paper evaluate on?
STCrowd, SiT and JRDB (multi-modal sets with LiDAR and egocentric images; evaluated on validation sets because test labels are withheld, using awareness labels the authors annotated themselves), plus the ETH/UCY benchmark, where only the reduced variant 'Mam' (no map or awareness input) is evaluated — and Table V omits the ETH scene itself.
Does the paper release code or the awareness annotations?
No code or data release statement appears in the paper's text. Reproducing it requires the authors' awareness-state annotations, whose annotation protocol (criteria, annotators, agreement) is not described, and no automatic extractor from egocentric images is presented even though the conclusion claims awareness states are 'extracted from egocentric images'.
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.
- Figure interiors (Figs. 1–5) are images not recoverable from text extraction; assessments of the qualitative results and of the Fig. 4 decay analysis rely on captions and on numbers quoted in the body text.
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.