Multipath: Multiple Probabilistic Anchor Trajectory Hypotheses for Behavior Prediction

Multipath:

In Multipath, we focus on the problem of predicting future agent states, which is the crucial task for robot planning in real-world environments. We are specifically interested in addressing this problem for self-driving vehicles, an application with a potentially enormous societal impact. Mainly, predicting the future of other agents in this domain is vital for safe, comfortable, and efficient operation.

E.g., it is important to know whether to yield to the vehicle if they are going to cut in front of our robot or when would be the best time to add into traffic. Such future prediction requires an understanding of a static and dynamic world context: road semantics (like lane connectivity, stop lines), traffic light information, and past observations of other agents, as in below Fig. A fundamental aspect of the future state prediction is that it is inherently stochastic, as agents can’t know each other’s motivations. When we are driving, we can never really be sure what other drivers will do next, and it is essential to consider multiple outcomes and their likelihood.

We seek the model of the future that can provide both:

(i) a weighted, parsimonious set of discrete trajectories that covers a space of likely outcomes and (ii) a closed-form evaluation of the likelihood of any trajectory. These two attributes enable efficient reasoning in relevant planning use-cases, e.g., human-like reactions to discrete trajectory hypotheses (e.g., yielding, following), and probabilistic queries such as the expected risk of collision in a space-time region. This model addresses these issues with critical insight: it employs a fixed set of trajectory anchors as the basis of our modeling. This lets us factor stochastic uncertainty hierarchically: First, intent uncertainty captures the uncertainty of what an agent intends to do and is encoded as a distribution over the set of anchor trajectories. Second, given an intent, control uncertainty represents our uncertainty over how they might achieve it. We assume control uncertainty is normally distributed at each future time step [Thrun05], parameterized such that the mean corresponds to a context-specific offset from the anchor state, with the associated covariance capturing the unimodal aleatoric uncertainty [Kendall17]. Fig. Illustrates a typical scenario where there are three likely intents given the scene context, with control mean offset refinements respecting road geometry, and control uncertainty intuitively growing over time. Our trajectory anchors are modes found in our training data in state-sequence space via unsupervised learning. These anchors provide templates for coarse-granularity futures for an agent and might correspond to semantic concepts like “change lanes,” or “slow down” (although to be clear, we don’t use any semantic concepts in our modeling).

Our complete model predicts a Gaussian mixture model (GMM) at each time step, with the mixture weights (intent distribution) fixed over time. Given such a parametric distribution model, we can directly evaluate the likelihood of any future trajectory and have a simple way to obtain a compact, diverse weighted set of trajectory samples: the MAP sample from each anchor-intent.