NVIDIA Unveils ENPIRE Robot Policy Framework
NVIDIA's GEAR Lab, Carnegie Mellon University, and UC Berkeley have developed ENPIRE, an autonomous closed-loop framework that enables AI agents to design, execute, and improve physical robot manipulation policies. By autonomously running trials, analyzing logs, and rewriting policy code, the system achieves high success rates on complex real-world tasks like GPU installation.
By automating the physical evaluation and feedback loop of robotic training, NVIDIA is shifting the paradigm from manual algorithmic tuning to agent-led autonomous discovery. While virtual simulation has advanced rapidly, real-world deployment has always been bottlenecked by human intervention; ENPIRE bridges this gap by proving that coding agents can manage physical hardware labs.
- –**Closed-Loop Autonomy:** Integrates automatic environment resetting and evaluation with agentic code generation to automate the entire development cycle.
- –**Tackling Real-World Complexity:** Demonstrated success on high-dexterity real-world tasks like GPU installation and zip-tie fastening.
- –**Hardware Parallelization:** Features multi-robot rollout capability, allowing coding agents to scale testing across physical workspaces.
DISCOVERED
2d ago
2026-06-17
PUBLISHED
2d ago
2026-06-17
RELEVANCE
AUTHOR
skyrain888