Time to Collision (TTC)

Time to Collision (TTC) is the estimated time remaining before two objects on a constant relative velocity course will collide if no evasive action is taken. It is calculated as the current separation distance divided by the closing speed (the magnitude of the relative velocity vector along the line of sight).

• Formula: TTC = Distance / |Relative Velocity|
• Assumption: Constant velocity (no acceleration).
• Output: A scalar value in seconds. A lower TTC indicates higher immediate risk.

The standard TTC calculation assumes both the agent and the obstacle maintain their current velocities. This is a kinematic simplification that makes TTC computationally lightweight and suitable for real-time systems.

• Limitation: It cannot account for accelerating, decelerating, or maneuvering obstacles.
• Application: Best used for short-term, tactical risk assessment where prediction horizons are brief (e.g., < 2-3 seconds).
• Extension: More complex models, like Trajectory Prediction, are required for longer horizons or non-linear motion.

TTC is rarely used in isolation. It is combined with other spatial and probabilistic metrics to form a comprehensive Collision Risk Assessment.

• Distance to Closest Point of Approach (DCPA): The predicted minimum separation distance.
• Probability of Collision: Often derived from distributions around TTC and DCPA estimates.
• Safety Margin: A buffer (e.g., TTC threshold of 3-5 seconds) added to account for system latency and uncertainty.
• Context: Priority of agents, zone rules, and operational state can adjust the risk interpretation of a given TTC value.

TTC operates at different levels of a safety stack. It triggers specific protocols based on threshold breaches, enabling layered responses.

• Warning (High TTC): Alerts a human operator or high-level planner (e.g., TTC > 8s).
• Tactical Avoidance (Medium TTC): Triggers reactive algorithms like Velocity Obstacle (VO) or Model Predictive Control (MPC) to compute a new path (e.g., 3s < TTC < 8s).
• Emergency Maneuver (Low TTC): Invokes a deterministic Emergency Stop Protocol or maximal deceleration command (e.g., TTC < 2s).

TTC is only as accurate as the pose and velocity estimates of the agent and obstacles. This creates a direct dependency on upstream perception systems.

• Required Inputs: Accurate position, velocity, and sometimes heading for both entities.
• Sensor Fusion: LiDAR, radar, and cameras fused together provide the most robust state estimation.
• Error Propagation: Sensor noise and latency directly affect TTC calculation, necessitating probabilistic representations and conservative safety margins.
• V2X Enhancement: Vehicle-to-Everything (V2X) communication can provide highly accurate intent and state data from cooperative obstacles, improving TTC reliability.

While foundational, the basic TTC model has known limitations that advanced methods address.

• Non-Linear Motion: Does not model turning or acceleration. Extended by trajectory prediction models.
• Multi-Agent Scenarios: Basic TTC is pairwise. Reciprocal Velocity Obstacle (RVO) algorithms use TTC-like concepts within a velocity-space framework for cooperative, multi-agent avoidance.
• Uncertainty: Probabilistic TTC (PTTC) models state estimates as distributions, yielding a probability distribution over possible collision times.
• Shape & Orientation: Ignores object geometry. Combined with Collision Cones or Occupancy Grids to account for size and orientation.

In warehouse and logistics automation, TTC is used for decentralized collision avoidance among Autonomous Mobile Robots (AMRs). Each robot calculates TTC with nearby agents and obstacles to make local, cooperative decisions—slowing down, yielding, or replanning paths. This enables safe, high-density operation without central micromanagement. Integration with the Velocity Obstacle (VO) family of algorithms allows robots to select velocities that maintain a positive, safe TTC.

For robots and autonomous vehicles operating near humans, TTC informs socially-aware navigation. By estimating TTC with pedestrians, systems can exhibit predictable, courteous behavior—such as slowing to allow crossing—that aligns with human expectations. This moves beyond hard collision avoidance to fluid co-existence. Research in this area often integrates TTC with models like the Social Force Model to generate more natural agent trajectories.

Within a heterogeneous fleet, TTC is a dynamic input for priority-based routing and exception handling. A high-priority agent (e.g., carrying a critical part) may be assigned a route that minimizes its TTC with all obstacles, while lower-priority agents yield. The orchestration middleware can use real-time TTC streams from all agents to detect potential deadlock scenarios and command pre-emptive resolutions before agents come to a full stop.

The Closest Point of Approach (CPA) is a fundamental pair of navigational metrics that predict the future separation between two moving objects. It consists of:

• Distance to CPA (DCPA): The predicted minimum distance between the objects.
• Time to CPA (TCPA): The time remaining until that minimum distance is reached. While Time to Collision (TTC) assumes a collision course (DCPA = 0), CPA provides a more general risk assessment. A small DCPA combined with a small TCPA indicates high collision risk, even if a direct impact is not certain. This is critical for maritime and aerial navigation where maintaining safe separation distances is the primary goal.

The Velocity Obstacle (VO) is a geometric algorithm that directly operationalizes collision prediction. For a given agent, it constructs a collision cone in velocity space—the set of all velocities that would result in a collision with a dynamic obstacle within a specified time horizon (often derived from TTC). The agent must then select a velocity outside this cone to avoid collision. This method provides a clear, computable link between an agent's chosen velocity and its future collision state, forming the basis for reactive, real-time avoidance in multi-agent systems like robotic fleets.

Collision Risk Assessment is the holistic process of evaluating the likelihood and severity of a potential collision. It synthesizes quantitative metrics like Time to Collision (TTC) and Distance to CPA (DCPA) with contextual factors such as:

• Agent and obstacle types and sizes.
• Environmental conditions (e.g., visibility, friction).
• System uncertainties (sensor noise, actuation lag).
• Operational rules (right-of-way, zones). The output is often a normalized risk score (e.g., 0 to 1) or a categorical level (Low, Medium, High, Critical) that informs the urgency and type of evasive action required by a Collision Avoidance System (CAS).

Model Predictive Control (MPC) is an optimization-based strategy that repeatedly solves a finite-horizon optimal control problem. For collision avoidance, the cost function minimizes deviation from a goal while constraints explicitly enforce that predicted future states avoid collisions, often using TTC or Velocity Obstacle conditions. MPC computes a short sequence of optimal control inputs (e.g., steering, acceleration), applies the first step, and then re-plans at the next time interval. This allows it to proactively shape a smooth, kinodynamically feasible trajectory that avoids obstacles, making it a powerful method for high-speed autonomous vehicles.

Trajectory Prediction is the process of forecasting the future states (position, velocity, acceleration) of dynamic obstacles. Accurate prediction is a prerequisite for calculating meaningful Time to Collision (TTC) and other risk metrics. Methods range from simple constant velocity or constant acceleration kinematic models to complex machine learning models (e.g., LSTMs, Graph Neural Networks) that learn patterns from historical data. For heterogeneous fleets, predicting the intent of human-operated vehicles or other agents with non-linear behavior is a major challenge that directly impacts the reliability of TTC-based safety systems.

Runtime Assurance (RTA) is a safety architecture designed to guarantee safe operation even when the primary controller (which may use complex, learning-based algorithms for planning) is untrusted. In collision avoidance, a simple, verifiable safety monitor continuously checks the primary controller's commands against a safety property—such as maintaining a positive Time to Collision (TTC) or staying outside a Velocity Obstacle cone. If a violation is predicted, the RTA system overrides the primary command with a safe alternative from a safety controller. This provides a formal safety layer for systems where the core AI cannot be fully certified.