Traditional trajectory forecasting models treat moving agents—whether they are pedestrians on a sidewalk or vehicles at an intersection—as passive particles governed by local physics. To predict where an agent will be a few seconds from now, standard architectures take a short snippet of past coordinates and extrapolate them step-by-step into the future.
As Monti et al. observe in DAG-Net: Double Attentive Graph Neural Network for Trajectory Forecasting, this paradigm effectively "pushes" the agent forward from the past based on immediate momentum. However, human navigation is fundamentally goal-oriented. We do not simply drift forward; we are "pulled" toward specific destinations.
By failing to account for this long-term intent, traditional autoregressive models suffer from two catastrophic flaws: compounding drift and modality collapse. Goal-conditioned motion forecasting flips this paradigm by breaking trajectory prediction into a two-stage hierarchical process: first estimating where an agent intends to go, and then routing how they will get there.
The Failure Modes of Pure Extrapolation
To understand why goal-conditioning works, we must first look at why step-by-step sequence generation fails over long horizons.
When a model predicts a trajectory autoregressively (generating waypoint $t+1$ based on waypoint $t$), small errors accumulate rapidly. A minor directional misalignment in the first few frames compounds over time, causing the predicted path to drift away from reality as the forecasting horizon extends. As highlighted by Camporese et al. in Distilling Knowledge for Short-to-Long Term Trajectory Prediction, this unconstrained sequence generation lacks a global anchor to keep long-horizon predictions physically and logically sound.
Furthermore, human movement is highly multimodal. An driver approaching an intersection might turn left, turn right, or continue straight. When a standard regression model attempts to minimize mean squared error across these distinct possibilities, it often suffers from modality collapse—producing a single, blurry path that averages out all options, resulting in an unrealistic trajectory down the middle of the intersection (Sun et al. in Generalizing Motion Planners with Mixture of Experts for Autonomous Driving).
Pure Extrapolation (The "Push" Approach):
[Past Trajectory] ---> [Step 1] ---> [Step 2] ---> [Step 3] ---> (Compounding Drift)
Goal-Conditioned (The "Pull" Approach):
[Past Trajectory] ---------------------------------------------> [Predicted Goal]
\ /
---> [Waypoint 1] ---> [Waypoint 2] --
Decomposing and Factorizing Uncertainty
Goal-conditioned forecasting resolves these issues by reframing navigation as a targeted reaching problem. As proposed by Dendorfer et al. in Goal-GAN: Multimodal Trajectory Prediction Based on Goal Position Estimation, explicitly estimating a long-term destination before generating a path simplifies the learning space.
Mathematically, this architecture works because it factorizes the joint probability distribution of a future trajectory into two distinct components. As detailed by Mangalam et al. in From Goals, Waypoints & Paths To Long Term Human Trajectory Forecasting (Y-net), trajectory uncertainty can be split into:
- Epistemic Uncertainty (Intent): Where is the agent going? This captures the high-level, discrete choices available to the agent.
- Aleatoric Uncertainty (Execution): What exact path will they take to get there? This captures the low-level, continuous variations in execution, such as swaying or adjusting speed.
By predicting a spatial probability distribution over potential endpoints first—such as a 2D heatmap or a discrete set of coordinate targets—the model isolates the primary source of multimodal ambiguity. Once a specific goal is sampled from this distribution, the downstream routing module can treat the destination as a fixed boundary condition.
As shown in PECNet (Mangalam et al. in It Is Not the Journey But the Destination: Endpoint Conditioned Trajectory Prediction), conditioning the path decoder on a fixed endpoint forces the generated waypoints to remain physically consistent with a plausible final location, effectively eliminating the risk of modality collapse.
Why Destination Anchoring Eliminates Drift
The presence of a fixed destination target acts as a global correction mechanism for the trajectory generator. In real-world environments, agents frequently deviate from their optimal path to avoid immediate hazards, such as a pedestrian stepping around a puddle or a car dodging a sudden obstacle.
A pure momentum model struggles with these deviations; it interprets the sudden evasive maneuver as a permanent change in direction, causing its long-term predictions to veer off-course. Conversely, a goal-conditioned model understands that local adjustments are temporary. This insight dates back to early social motion modeling by Pellegrini et al. in You'll never walk alone: Modeling social behavior for multi-target tracking, which demonstrated that goal-oriented agents naturally return to their primary path after avoiding collisions.
By utilizing a dual-channel architecture—where one channel maintains a stable destination target and a secondary channel handles local path adjustments (Tran et al. in Goal-driven Long-Term Trajectory Prediction)—the forecasting system decouples global intent from local dynamics. Even if the intermediate waypoints are noisy or perturbed, the trajectory is guaranteed to terminate at a logical, reachable endpoint, drastically reducing error accumulation over extended time horizons.
Grounding Intent in Environmental Geometry
In structured environments like urban roadways, the space of plausible goals is heavily constrained by geometry. Goal-conditioned frameworks exploit this by grounding their first-stage target estimation directly in map features.
Instead of predicting arbitrary coordinates in a vacuum, target-driven architectures like TNT, surveyed by Liu et al. in A Survey on Deep-Learning Approaches for Vehicle Trajectory Prediction in Autonomous Driving, project candidate targets onto discrete map elements, such as lane polylines or crosswalk zones.
Map-Grounded Target Selection:
[Intersection Map] ---> Sample Candidate Targets (Lane Polylines)
---> Score Targets based on Past Motion
---> Interpolate Feasible Paths to Top Targets
This map-grounding process ensures that the model's intent-recognition layer is structurally aligned with traffic rules and physical boundaries before any continuous path interpolation begins. By converting an unconstrained continuous search space into a discrete selection problem over valid environmental targets, goal-conditioned motion forecasting achieves both the robustness and the multi-modal flexibility required for safe autonomous navigation.
Key Foundational Work
Learning and Inferring Transportation Routines (2004) — contributed 2 concept elements to this direction.