arXiv:2603.22435v2 Announce Type: replace-cross Abstract: "Code-as-Policy" considers how executable code can complement data-intensive Vision-Language-Action (VLA) methods, yet their effectiveness as autonomous controllers for embodied manipulation remains underexplored. We present CaP-X, an open-access framework for systematically studying Code-as-Policy agents in robot manipulation. At its core is CaP-Gym, an interactive environment in which agents control robots by synthesizing and executing programs that compose perception and control primitives. Building on this foundation, CaP-Bench eval
The concept of 'Code-as-Policy' for embodied AI is gaining traction as researchers seek more robust and interpretable control methods for robotic systems, moving beyond purely data-driven approaches.
This framework offers a standardized way to benchmark and improve coding agents in robot manipulation, which is critical for advancing the reliability and capabilities of autonomous robotic systems.
The development of open-access benchmarking tools like CaP-X and CaP-Gym provides a structured pathway for evaluating and iterating on AI agents that control robots through generated code.
- · Robotics research institutions
- · AI agent developers
- · Automation industries
- · Companies relying solely on black-box VLA models for critical robot control
Improved performance and broader adoption of AI agents for complex robot manipulation tasks in commercial and industrial settings.
Increased demand for sophisticated code generation and verification tools specifically tailored for robotic applications.
Accelerated development of general-purpose robots capable of adapting to diverse environments through on-the-fly code synthesis and execution.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.AI