ScheduleStream: Temporal Planning with Samplers for GPU-Accelerated Multi-Arm Task and Motion Planning & Scheduling

ScheduleStream introduces a GPU-accelerated framework for Task and Motion Planning & Scheduling (TAMPAS) that enables bimanual and humanoid robots to coordinate parallel arm movements efficiently. The system models temporal dynamics through hybrid durative actions and produces more optimized schedules than traditional TAMP algorithms that typically move one arm at a time.

ScheduleStream addresses a fundamental computational bottleneck in multi-arm robotics by extending Task and Motion Planning (TAMP) algorithms to handle temporal scheduling alongside spatial planning. Traditional TAMP systems struggle with bimanual and humanoid robots because they treat arm movements sequentially, wasting efficiency gains that parallel coordination could provide. The hybrid discrete-continuous action space grows exponentially with each additional arm, making naive approaches computationally intractable.

This work builds on decades of robotics research in planning and scheduling, but represents a meaningful convergence by treating durative actions as first-class primitives rather than afterthoughts. The introduction of GPU-accelerated sampling within the planning framework capitalizes on modern hardware capabilities, enabling faster exploration of the expanded action space. By modeling actions that can start asynchronously and persist for parameter-dependent durations, ScheduleStream captures realistic robot behaviors without domain-specific engineering.

The practical implications extend across manufacturing, logistics, and autonomous systems. Real-world bimanual tasks—from assembly to manipulation—become more tractable and efficient when arms can work in parallel rather than serial sequences. This reduces cycle times and increases throughput, directly impacting productivity metrics that drive adoption. For robotics developers and researchers, the framework's domain-independent algorithms reduce engineering overhead when adapting to new tasks or robot morphologies.

Looking ahead, the critical evaluation metric will be real-world deployment success and scalability to more complex scenarios with longer planning horizons. The field should monitor whether GPU-accelerated sampling remains competitive as problem complexity grows beyond current demonstrations, and whether the approach generalizes to robots with more than two arms or hybrid legged-armed systems.

• →ScheduleStream enables parallel multi-arm robot motion planning where previous TAMP systems produced sequential arm movements
• →GPU-accelerated sampling within the framework significantly speeds up planning for hybrid discrete-continuous action spaces
• →Hybrid durative actions model realistic temporal dynamics with asynchronous starts and parameter-dependent durations
• →Domain-independent algorithms eliminate application-specific engineering while improving solution efficiency
• →Real-world bimanual robot demonstrations validate the framework's practical utility for manufacturing and manipulation tasks