💻 computer science

Social-Mamba: Socially-Aware Trajectory Forecasting with State-Space Models

Social-Mamba is a novel trajectory forecasting architecture that leverages Selective State-Space Models and a Cycle Mamba block to efficiently model social interactions with linear computational complexity, achieving state-of-the-art accuracy and scalability across multiple benchmarks.

Original authors: Po-Chien Luan, Wuyang Li, Yang Gao, Alexandre Alahi

Published 2026-05-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Po-Chien Luan, Wuyang Li, Yang Gao, Alexandre Alahi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are standing in the middle of a very crowded dance floor. You need to predict where everyone else will move in the next few seconds so you don't bump into them. This is exactly what computers try to do for robots and self-driving cars, a task called trajectory forecasting.

The paper introduces a new AI model called Social-Mamba that is exceptionally good at this dance-floor prediction, but it does so in a much smarter and faster way than previous methods.

Here is how it works, broken down with simple analogies:

1. The Problem: The "Too Many Neighbors" Dilemma

Previous AI models (like Transformers) tried to look at every person on the dance floor and calculate how they interact with every other person simultaneously.

The Analogy: Imagine trying to have a conversation with 100 people at once, where you have to listen to every single pair of people talking to each other. As the crowd grows, the amount of work explodes. It gets so heavy and slow that the computer gets overwhelmed.
The Old Fix: Some newer models tried to use a "one-way street" approach (called State-Space Models or Mamba) to save energy. They looked at people one by one in a line.
The Flaw: Social interactions aren't a straight line; they are a messy, 3D web. Forcing people into a single line destroys the spatial context (who is actually standing next to whom). Also, looking only forward means you miss the fact that what someone does now might be explained by what they did a moment ago (a "retrospective" view).

2. The Solution: Social-Mamba's "Smart Grid"

Social-Mamba solves this by organizing the chaos into a structured, efficient process.

Step A: The "Ego-Centric" Grid
Instead of looking at the whole room as a giant mess, the model focuses on You (the "Ego"). It draws an invisible circle around you and organizes everyone inside that circle into a neat grid based on where they are relative to you.

Analogy: Instead of trying to memorize the whole city map, you just look at the street corner you are standing on and organize the people around you by "left," "right," "front," and "back."

Step B: The "Social Triplet" (Three Ways of Thinking)
The model doesn't just look at the grid once. It breaks the interaction down into three specific "scans" or ways of thinking, like a detective looking at a crime scene from three angles:

Time Scan: How does this specific person move over time? (Are they speeding up? Stopping?)
Ego Scan: How does this person affect you right now? (Are they blocking your path?)
Goal Scan: How does this person affect where you are trying to go? (Are they cutting off your route to the exit?)

Step C: The "Cycle Mamba" (The Magic Loop)
This is the paper's biggest innovation. Standard AI models usually look forward (past to future) or backward (future to past) separately. Social-Mamba uses a Cycle Mamba block.

The Analogy: Imagine reading a book. A normal model reads from page 1 to 100, then starts over and reads 100 to 1. They are two separate readings.
Social-Mamba reads page 1 to 100, but the memory of the last page (100) is immediately carried over to help understand page 1. It creates a continuous loop where the "future" context helps explain the "past" actions instantly. This allows the model to understand that a person's sudden turn was actually a reaction to something that happened a split second earlier.

3. The Results: Faster and Smarter

The authors tested Social-Mamba on five different datasets, including:

NBA: Tracking basketball players.
SDD: Pedestrians on a university campus.
JRDB: People walking around a robot in a city.

The Claims:

Accuracy: It predicts where people will go more accurately than the current best models (State-of-the-Art).
Efficiency: It uses 75% fewer parameters (think of this as the model's "brain size" or memory) and requires 54% less computing power to run.
Speed: It can make predictions in about 3.4 milliseconds, which is fast enough for real-time robots.

4. Why It Matters (According to the Paper)

The paper claims this is the first model to successfully use the efficient "Mamba" architecture for complex social interactions without losing the ability to understand 2D space. It proves that you don't need a massive, slow computer to understand human crowds; you just need a smarter way to organize the data.

The authors also showed that this model is flexible enough to be swapped into other advanced systems (like "flow-matching" frameworks) to make those systems faster and more accurate too.

In short: Social-Mamba is like a super-efficient dance partner that can predict the moves of a whole crowd by organizing them neatly around you and using a "looping" memory to understand the past and future simultaneously, all while using a fraction of the computer power required by older models.

Technical Summary: Social-Mamba

Problem Statement

Human trajectory forecasting is critical for safe navigation in crowded environments, such as autonomous robotics and sports analytics. The core challenge lies in balancing predictive accuracy with computational efficiency. While recent state-of-the-art methods rely on Transformer-based attention mechanisms to capture complex social interactions, these approaches incur quadratic computational costs that scale poorly with the number of agents. Conversely, Selective State-Space Models (SSMs), specifically Mamba, offer linear-time complexity and are effective for long-range dependencies. However, applying SSMs directly to social interactions presents two fundamental conflicts:

Structural Mismatch: SSMs are designed for ordered, one-dimensional sequences, whereas social interactions occur in unstructured 2D or 3D spaces. Forcing agents into arbitrary sequences risks destroying spatial information.
Information Flow Mismatch: Standard SSMs process information unidirectionally. Social interactions, however, are often holistic and bidirectional; the intent of an early action is often clarified by subsequent events, a context that purely forward-pass models cannot capture.

Prior attempts to use Mamba for trajectory prediction have largely restricted the model to temporal history only or modified its core mechanism to resemble attention, thereby failing to leverage the native efficiency of SSMs for social reasoning.

Methodology

The authors propose Social-Mamba, a forecasting architecture that reformulates social interactions as structured sequential processes to align with the native design of SSMs. The framework consists of three primary components:

1. Cycle Mamba (CM) Block

To address the bidirectional nature of social reasoning, the authors introduce the Cycle Mamba block. Unlike conventional bidirectional models that process forward and backward passes independently and fuse them only at the output, the CM block creates a continuous state flow.

Mechanism: It constructs a cycle sequence by concatenating the reversed input sequence with the original forward sequence ( $S_{cycle} = [S_{bwd}; S_{fwd}]$ ).
State Propagation: A single Mamba model processes this unified sequence. Crucially, the forward pass is initialized with the final hidden state of the backward pass. This ensures that the forward processing of an agent's history is explicitly conditioned on a compressed representation of the future context, enabling seamless, integrated bidirectional fusion with fewer parameters.

2. Social Triplet Factorization

To resolve the structural mismatch between 1D SSMs and 2D social scenes, the authors organize agents on an ego-centric social grid and decompose interactions into three parallel sequential scans:

Temporal Interaction: Processes individual agent histories independently to capture motion dynamics.
Egocentric Interaction: Inserts the ego agent's current state token into each neighbor's sequence. This allows the model to scan how neighbors influence the ego's current state.
Goal-Centric Interaction: Appends the ego agent's goal token to neighbor sequences to model how neighbors affect the ego's path toward its destination.

In both egocentric and goal-centric scans, the ego's state or goal tokens are inserted into neighbor sequences before the Cycle Mamba scan, allowing different interaction facets to be processed in a temporal manner.

3. Social Fusion and Decoding

The outputs of the three interaction streams ( $Z^T, Z^E, Z^G$ ) are fused using a learnable social gate. A gating network computes dynamic weights for each interaction type, allowing the model to adaptively emphasize relevant contexts (e.g., downweighting social influence when the ego agent moves independently). Finally, a global scan is performed along the agent axis to aggregate information, followed by a decoder (a bidirectional Mamba and MLP heads) to predict $K$ future trajectories.

Key Contributions

First Native SSM for Social Reasoning: Social-Mamba is the first trajectory forecasting architecture built entirely on the Mamba framework that addresses social interactions directly, rather than restricting SSMs to temporal modeling alone.
Cycle Mamba Block: A novel module enabling continuous bidirectional information flow, which improves contextual understanding and reduces parameter count by half compared to standard bidirectional designs.
Structured Social Scanning: The introduction of ego-centric grids and social triplet factorization successfully maps unstructured 2D social dynamics onto the 1D sequential processing of SSMs without sacrificing spatial information.
Efficiency and Scalability: The architecture achieves linear-time complexity, offering superior computational scalability compared to quadratic-cost attention mechanisms.

Experimental Results

The authors evaluated Social-Mamba on five benchmark datasets: NBA (Full, Scoring, and Rebounding splits), Stanford Drone Dataset (SDD), and JRDB.

Accuracy: Social-Mamba achieved state-of-the-art performance across all benchmarks. On the NBA-Full dataset, it reduced Average Displacement Error (ADE) to 0.72m and Final Displacement Error (FDE) to 0.92m, outperforming strong baselines like Multi-Transmotion and OmniTraj.
Efficiency: The model demonstrated significant gains in parameter efficiency and computational cost. Compared to prior state-of-the-art methods, Social-Mamba reduced parameters by 75% and GFLOPs by 54%.
Flexibility: When integrated as a drop-in replacement for the Transformer encoder in a flow-matching framework, Social-Mamba improved accuracy while using an encoder 2.3× smaller than the Transformer-based counterpart.
Qualitative Analysis: Visualizations confirmed the model's ability to handle abrupt turns, non-linear behaviors, and dense crowds, generating diverse and plausible multimodal trajectories.

Significance and Claims

The paper claims that Social-Mamba establishes a new foundation for trajectory forecasting by successfully bridging the gap between the efficiency of State-Space Models and the complexity of unstructured social interactions. By demonstrating that SSMs can be adapted to model social dynamics without resorting to attention-like mechanisms or sacrificing bidirectional reasoning, the authors position Social-Mamba as a flexible, robust, and highly efficient alternative to Transformer-based paradigms. The work suggests that the inherent linear scalability of SSMs makes them particularly well-suited for real-time applications in crowded, multi-agent environments where computational resources are constrained.