Robotics
New robot planning framework combines black-box skills for complex multi-step tasks
Researchers demonstrate how heterogeneous robot skills—from learned policies to force controllers—can be integrated into hierarchical planning systems for real-world manipulation problems.
Robotic systems have long struggled to plan long-horizon tasks that require sequencing multiple capabilities. A new approach called Task and Skill Planning (TASP) addresses this challenge by allowing robots to leverage diverse, pre-built skills—including learned models, force-controlled actions, and opaque policies—within a unified hierarchical planning framework.
The Challenge of Skill Integration#
Traditional task and motion planning (TAMP) systems assume that high-level robot actions map cleanly to kinematic motion plans. However, modern robotic applications increasingly rely on learned controllers, closed-loop force feedback, and vendor-supplied proprietary policies that function as black boxes. Integrating these heterogeneous components into planning without losing the object-centric reasoning that makes TAMP solvers effective has remained an open problem.
How TASP Works#
The new method preserves the interpretable failure reasoning characteristic of TAMP while bridging consecutive skills using Composable Interaction Primitives (CIPs). These primitives generate head and tail motion sequences that connect the end state of one skill to the starting conditions of the next, enabling seamless transitions between actions. This approach supports both planning-time refinement—optimizing the overall trajectory before execution—and execution-time adjustment, allowing the system to respond to real-world variations.
Real-World Validation#
The team demonstrated TASP on both a bimanual manipulator and a mobile manipulator, testing the approach on multi-room mobile manipulation tasks with non-monotonic structure (scenarios where goal states cannot be achieved through a simple forward sequence). The framework's ability to synthesize bridging motions allowed diverse robot platforms to combine heterogeneous skills effectively, solving complex, long-horizon problems that would be difficult to address with conventional planning alone.
The research opens new possibilities for deploying pre-existing and learned skills across different robotic platforms without requiring ground-up replanning.
