Core Concepts
Our framework enables the generation of diverse and natural pedestrian animations that adhere to given trajectories and motion content, enhancing the realism and adaptability of pedestrian simulations for autonomous vehicle training.
Abstract
The paper presents PACER+, a simulation-based framework for generating diverse and natural pedestrian animations on-demand for autonomous vehicle (AV) simulation systems. The key contributions are:
The framework combines motion tracking and trajectory following tasks through a joint training scheme, enabling fine-grained control over different body parts while ensuring smooth motion, terrain compatibility, and adherence to the provided trajectory.
PACER+ supports the generation of diverse pedestrian behaviors from various sources, including generative models, pre-captured motions, and videos, in both manually built and real-world scanned environments.
The framework achieves zero-shot recreation of real-world pedestrian animations into simulation environments, automatically infilling missing parts.
The core insight is the synergy between motion imitation and trajectory following tasks. The lower-body motion is influenced by the trajectory and terrain, while the upper-body motion has the flexibility to encompass a diverse range of motions. The framework employs a per-joint spatial-temporal mask to indicate the presence of reference motion for the policy to track, enabling the concurrent tracking of trajectory and imitation of reference motion.
The evaluation demonstrates that PACER+ outperforms the state-of-the-art PACER framework in terms of motion quality and diversity, while also achieving superior motion tracking performance on both synthetic and real-world scenarios.
Stats
The paper presents the following key metrics and figures:
Motion Fréchet Inception Distance (FID) and Diversity metrics to evaluate the quality and diversity of synthesized animations.
Mean Per-Joint Position Error (Empjpe) and Global Mean Per-Joint Position Error (Egmpjpe) to evaluate tracking accuracy between the simulated character and reference motion.
Foot sliding (FS) and foot penetration (FL) metrics to evaluate the physical attributes of the animation.
Velocity (Vel) and acceleration (Accel) metrics to measure motion jitter.
Quotes
"Our framework offers richer zero-shot control beyond trajectory following and enables the creation of diverse animation in both manual and real-world scenarios, to meet the demand for more controllable generation."
"The key insight behind PACER+ lies in the synergy between motion imitation and trajectory following tasks."
"Notably, our framework achieves the zero-shot recreation of real-world pedestrian animations into simulation environments, where the missing part will be infilled automatically."