Fast Online Learning of CLiFF-maps in Changing Environments

TL;DR

The paper tackles non-stationary human motion in dynamic environments by online updating the CLiFF-map, a probabilistic representation of local flow patterns. It introduces per-location SWGMMs updated with a stochastic EM procedure, using dynamic step sizes $\gamma_k = N_k / N_{\mathrm{ind}}^k$ and a decay rule $N_{\mathrm{ind}}^k = \lambda N_{\mathrm{ind}}^{k-1} + N_k$ to balance history and new data, with potential component splitting guided by $\eta_{\mathrm{thres}}$. Experimental results on synthetic den520d and real ATC data show that the online method achieves lower NLL and drastically faster runtimes than history or interval baselines, while providing accurate and up-to-date flow representations that improve flow-aware planning. The work demonstrates practical benefits for real-time robotics in dynamic human environments and offers a memory-efficient alternative to re-learning MoDs from scratch.

Abstract

Maps of dynamics are effective representations of motion patterns learned from prior observations, with recent research demonstrating their ability to enhance various downstream tasks such as human-aware robot navigation, long-term human motion prediction, and robot localization. Current advancements have primarily concentrated on methods for learning maps of human flow in environments where the flow is static, i.e., not assumed to change over time. In this paper we propose an online update method of the CLiFF-map (an advanced map of dynamics type that models motion patterns as velocity and orientation mixtures) to actively detect and adapt to human flow changes. As new observations are collected, our goal is to update a CLiFF-map to effectively and accurately integrate them, while retaining relevant historic motion patterns. The proposed online update method maintains a probabilistic representation in each observed location, updating parameters by continuously tracking sufficient statistics. In experiments using both synthetic and real-world datasets, we show that our method is able to maintain accurate representations of human motion dynamics, contributing to high performance flow-compliant planning downstream tasks, while being orders of magnitude faster than the comparable baselines.

Figures (8)

• Figure 1: Online update results of toy example data compared with using only new observations and using all observations to build the model. The top row shows raw observations for each of the eight directions $(0^\circ, 45^\circ, 90^\circ, 135^\circ, 180^\circ, 225^\circ, 270^\circ, 315^\circ)$, provided in each iteration $k$. Blue arrows depict the mean vectors of a CLiFF Gaussian mixture model, with transparency indicating the component weights. Three modeling approaches are compared: the second row shows models built using only observations from the current iteration $k$; the third row shows models built with cumulative observations from iteration 1 to $k$, which are overgeneralized and fail to prioritize recent observations; the fourth row shows the proposed online update method, which incorporate new data while retaining relevant historical patterns, offering a dynamic representation of the motion pattern over time.
• Figure 2: The den520d synthetic dataset simulates trajectories between positions S1, S2, G1 and G2. The initial flow is shown in Condition A (left). To simulate the flow change, the start (S1, S2) and goal (G1, G2) positions are switched Condition B (right), reversing the dominant flow direction.
• Figure 3: Runtime of each iteration in the ATC (left) and den520d (right) datasets. In den520d, the first 10 batches are in Condition A and the second 10 batches are in Condition B. In both datasets, the online model shows significantly reduced runtime compared to the history and interval models.
• Figure 4: An example from the east corridor in the ATC dataset. Blue arrows show the mean vectors of SWGMMs, with transparency indicating component weights. The interval model (top row) uses only the last 60 minutes of data, disregarding previously learned patterns (e.g., at 2 PM and 3 PM). The history model (second row), which treats all observations equally, tends to obscure patterns and fails to capture dominant movements effectively. Conversely, the online model (third row) adapts more effectively to changing flows and accurately represents the primary patterns.
• Figure 5: Left: NLL evaluation results at different times of a day in the ATC dataset. The STeF-map struggles to accurately capture human motion flow, resulting in poor performance. Notably, the interval model exhibits worse NLL results mostly after 16:00. It struggles to capture the full spectrum of motion patterns and is not able to effectively manage outlier trajectories outside of the main flow. Right: NLL evaluation results over batches in Condition B of the den520d dataset. The online model quickly adapts to changes in human flow, achieving lower NLL values in fewer batches.