How Acceleration-Spectrum Flow Forecasts Future Motion for Groups of Moving Agents

How Acceleration-Spectrum Flow Forecasts Future Motion for Groups of Moving Agents

Ziyang Xu, Peng Han, Hai Zhou

7 min read•Nov 22, 2026

SpecFlow predicts multiple plausible paths for interacting people by generating future acceleration in a compact frequency-based representation, then reconstructing velocity and position from each person’s observed state. Across pedestrian, basketball, and aerial-view datasets, it ranks first on every ETH/UCY scene, leads four-second NBA average error, and ties the best final error—showing how a structured motion model can improve forecasts without scene imagery.

What Did Researchers Build?

SpecFlow is a system for forecasting the future paths of several moving agents at once. Given observed motion, it generates a set of possible futures rather than one fixed prediction. That matters because people can change direction, choose different destinations, or react to one another in ways that are hard to predict from a single path.

The central design choice is to generate future acceleration, rather than directly generate future positions. Acceleration describes how velocity changes; positions and velocities can then be reconstructed from it. SpecFlow represents the acceleration sequence using a discrete cosine transform, or DCT: a fixed mathematical basis that expresses a motion sequence as a blend of slow, broad changes and faster, more local changes. The model keeps only a selected portion of these components to control how much detail it represents.

To handle different possible destinations and interactions, SpecFlow also predicts endpoint proposals and uses them to guide its motion generator. Its tests cover pedestrian movement, basketball player motion, and aerial-view trajectories. The basketball examples forecast all ten players and the ball together, rather than treating each path as an independent prediction.

Overview of SpecFlow's acceleration-spectrum trajectory forecasting pipeline

What Were the Key Results?

SpecFlow’s reported results are strongest on ETH/UCY, a collection of pedestrian scenes: it ranks first in every scene under best-of-20 evaluation. That metric asks whether at least one of 20 generated paths comes close to the observed future. On NBA trajectories, SpecFlow records the lowest reported average displacement error (ADE) at four seconds and ties the best final displacement error (FDE). On SDD, an aerial-view trajectory dataset, it remains competitive.

BenchmarkADEFDEReported result
ETH/UCY0.150.22Ranks first in every scene
NBA, four seconds0.690.86Best ADE; tied best FDE
SDD7.3811.57Competitive result

ADE measures average distance between predicted and observed positions across the forecast; FDE measures the distance at the final forecast point. Lower is better. The paper reports these values using best-of-20 evaluation, so they describe how close the strongest sample comes—not whether all generated possibilities are accurate or well calibrated.

A controlled comparison shows that the representation contributes to the gains. Replacing frame-by-frame position generation with the full acceleration-spectrum model reduces ETH/UCY ADE from 0.18 to 0.15 and FDE from 0.28 to 0.22. The authors report smaller relative reductions on NBA and SDD, too. The half-spectrum setting performs best across all three benchmarks, while keeping only one quarter or all coefficients increases error.

How Does SpecFlow Work?

The model starts with the agents’ observed positions and velocities. For each agent, it represents the future acceleration sequence as DCT coefficients. DCT components are ordered from low to high temporal frequency: the early components capture broad, gradual changes, while later ones can represent sharper turns and other quicker changes. SpecFlow retains the first portion of this ordered representation, with the half-spectrum setting giving the best reported results.

A flow-matching network transforms Gaussian noise into plausible future acceleration coefficients. In plain terms, the network learns a path from random starting values to coefficients that resemble valid future motions. The process is conditioned on the observed context and on predicted endpoint proposals, which give the generator cues about where an agent might end up.

The resulting coefficients are converted back into an acceleration sequence. Two fixed cumulative-sum operations then reconstruct velocity and position from the last observed velocity and position. These operations are analytical: they are built into the model’s reconstruction step rather than separately learned. This keeps the generated motion tied to its starting state and gives the model a direct way to supervise position, velocity, and acceleration together.

SpecFlow’s attention mechanism handles two kinds of relationships in alternating steps. One step exchanges information among multiple candidate futures, helping the model represent different possible modes. The other exchanges information among agents, so each joint forecast can account for interaction. This design is especially relevant when several people or players move in the same scene.

Example of SpecFlow forecasting the coordinated motion of basketball players and the ball

Why Does This Matter for Robotics?

Robots moving around people need more than a guess about where each person will be next. They need to account for multiple possible destinations and for the way nearby agents influence one another. SpecFlow’s combination of multiple candidate futures, joint agent modeling, and explicit motion dynamics addresses those forecasting challenges in a structured way.

The acceleration-spectrum approach also offers a useful engineering trade-off. A compact representation can focus the model on broad movement patterns without discarding every detail of a turn or interaction. The experiments suggest that retaining too few components loses useful motion changes, while retaining every component can expose the model to high-frequency variation that contributes less to displacement accuracy. Choosing a middle bandwidth may therefore help balance smoothness and responsiveness.

However, these experiments evaluate human trajectory forecasting, not robot navigation or fleet operation. SpecFlow is not a complete planning or collision-avoidance system: a robot would still need to combine predictions with sensors, maps, safety constraints, and a control policy. For teams considering where motion prediction could eventually fit into deployments, warehouse robots offer one relevant hardware category—but this study does not test those robots.

What Are the Limitations and Open Questions?

SpecFlow uses motion history and endpoint predictions, but it does not incorporate scene imagery, walkable-area maps, or semantic cues about destinations. Those signals could matter when geometry alone does not reveal where someone is likely to go. The paper notes that the gains are smaller on SDD, but that benchmark does not isolate the effect of missing visual context, so it cannot establish the cause.

Best-of-20 ADE and FDE measure whether a set of samples contains a close match. They do not show whether the probabilities assigned to those futures are reliable, or whether the model’s full distribution is a good fit. The main experiments also use one training seed, so the reported results do not capture variation between independent training runs. Further testing would need calibration-aware measures, repeated runs, and controlled experiments with scene information.

Finally, strong trajectory forecasts do not by themselves demonstrate safe robot behavior. Deployment would require evaluation of how predictions affect planning and operation, including cases where agents move unexpectedly.

What Questions Do Readers Ask?

What does “acceleration spectrum” mean here? It means the model represents future acceleration as a set of frequency components, from broad changes to finer motion details, rather than generating each future position directly.

Does SpecFlow predict just one future path? No. It generates multiple candidate futures and uses endpoint proposals to help represent different possible destinations.

Was SpecFlow tested on robots? No. The reported benchmarks cover human trajectories, including pedestrians, basketball players, and aerial-view motion.

Does a strong best-of-20 score mean the forecasts are well calibrated? No. It shows that at least one sample can be close to the observed path, not that the model’s probabilities are reliable.

Conclusion

SpecFlow shows how generating acceleration in a compact frequency-based representation can support accurate multi-agent forecasts across several motion benchmarks. Its results are promising for trajectory prediction, while scene awareness, calibration, and robot deployment remain open steps.