In the world of autonomous systems, predicting the future is often a matter of drawing a straight line. Whether navigating a drone through a forest or a self-driving car through an intersection, many motion planning frameworks rely on the Constant Velocity Assumption (CVA). This assumption posits that dynamic agents—pedestrians, vehicles, or aircraft—will maintain their current speed and direction over a short planning horizon.
While computationally efficient, this linear simplification creates a fundamental bias. In reality, the world is non-linear, stochastic, and reactive. When agents accelerate, brake, or turn, the linear models that autonomous systems use to "see" the future begin to break down, leading to safety violations and inefficient behavior.
The Appeal of First-Order Dynamics
The CVA is ubiquitous because it simplifies the complex geometry of interaction into a manageable mathematical problem. As noted in P4P: Conflict-Aware Motion Prediction for Planning in Autonomous Driving, kinematic models based on constant velocity help reduce prediction "jitter" and provide a stable baseline for planning. In aviation, Inseok Hwang et al. observe that assuming a constant relative heading between vehicles simplifies trajectory projection, making mid-air conflict detection computationally feasible.
However, this stability comes at the cost of accuracy. By ignoring higher-order dynamics like acceleration and jerk, these models fail to capture the "intent" behind a maneuver. For instance, Hanwool Woo et al. highlight that CVA-based systems struggle to anticipate lane changes, as the initial lateral velocity of a vehicle beginning a slow merge is often too small to distinguish from straight-line driving.
Where Linearity Fails
The failure of the constant velocity assumption is most evident in "high-dynamic" scenarios—situations where agents rapidly change their state.
- Longitudinal Dynamics: In stop-and-go traffic, vehicles are rarely at a constant speed. Kun Jiang 0002 et al. argue that long-range risk assessment fails when speed variations from rapid acceleration or braking are ignored. Similarly, Kyungtae Han et al. note that distributed control strategies for connected vehicles degrade when they cannot account for the non-linear reference trajectories of a lead agent.
- Lateral and Angular Maneuvers: Linear extrapolation cannot predict a curve. David Wolinski et al. point out that human navigation in crowds involves constantly changing paths and speeds that linear models simply cannot replicate. In robotics, Dinesh Manocha et al. show that assuming constant velocity in dense, chaotic environments leads to collisions because agents frequently deviate from linear paths to avoid one another.
- Environmental Context: Agents often change velocity due to the environment, not just internal physics. Beihao Xia et al. note that trajectory models struggle when agents abruptly stop or start without an obvious external trigger, a behavior common in crowded plazas.
The Cost of the Bias
When a planner assumes a linear future that doesn't manifest, the consequences range from "freezing" to catastrophic failure.
In urban driving, Yan Ding et al. explain that safety estimation frameworks using CVA cannot accurately react to the sudden accelerations of human drivers. This often forces the autonomous vehicle into overly conservative behavior. Conversely, Ichiro Hasuo et al. demonstrate that while safety buffers might prevent a collision, the inability to predict dynamic maneuvers can cause a system to fail its high-level goals, such as failing to complete a merge because it couldn't anticipate a gap opening up.
In robotics, Zhuozhu Jian et al. find that constant velocity assumptions lead to momentary inaccuracies in safety buffers. If a pedestrian lunges or stops abruptly, a robot relying on a linear projection may find its "safe zone" suddenly occupied, as Yingjian Wang et al. observe in autonomous drone flights.
Moving Beyond Constant Velocity
To overcome this bias, researchers are moving toward models that treat velocity as a dynamic variable rather than a constant.
- Adaptive Safety Parameters: John M. Dolan et al. propose learning the specific safety parameters and driving styles of surrounding agents in real-time, allowing the vehicle to adapt to non-linear behaviors like emergency braking.
- Probabilistic and Generative Modeling: Instead of a single line, systems like those proposed by Brian Yang et al. use diffusion models to generate a distribution of realistic, non-linear trajectories that account for multi-agent interactions.
- Hybrid State Estimation: The RMIMM algorithm introduced by Inseok Hwang et al. uses multiple models to identify the "flight mode" of an aircraft, allowing the system to switch between different motion models as the vehicle begins to maneuver.
By acknowledging that the straight line is an exception rather than the rule, these approaches allow autonomous systems to navigate the inherent unpredictability of the real world. As Marco Hutter suggests, moving toward intent-based and non-linear motion prediction is the necessary next step for robots to truly share public spaces with humans.
Key Foundational Work
Object Detection and Tracking for Autonomous Navigation in Dynamic Environments (2010) — contributed 1 concept elements to this direction.
Flight-Mode-Based Aircraft Conflict Detection Using a Residual-Mean Interacting Multiple Model Algorithm (2003) — contributed 1 concept elements to this direction.