ENPIRE is a NVIDIA/CMU/UC Berkeley harness framework for agentic robot policy self-improvement in the real world. It turns robotic manipulation research into a loop that coding agents can operate: reset the physical scene, run a policy rollout, automatically verify the result, inspect logs/videos/traces, and edit policy or training code for the next iteration.source: nvidia-enpire-agentic-robot-policy-self-improvement-2026.md
The acronym names the four core modules: EN for the environment interface, including automatic reset, safety, verification, and logging; PI for policy improvement through heuristics, tool calling, behavior cloning, offline RL, or online RL; R for rollouts across one or more physical robots; and E for evolution, where agents compare branches, reuse successful recipes, and prune failed hypotheses. This makes ENPIRE a concrete robotics instance of harness-engineering and agent-loops.source: nvidia-enpire-agentic-robot-policy-self-improvement-2026.md
The showcased tasks include Push-T, pin insertion, GPU insertion, and zip-tie manipulation. The project reports policies reaching up to 99% pass@8 success across showcased manipulation tasks, with coding agents autonomously searching through algorithmic variants under real-world feedback. The page also evaluates Codex with GPT-5.5, Claude Code with Opus 4.7, and Kimi Code with Kimi K2.6 on AutoEnvBench.source: nvidia-enpire-agentic-robot-policy-self-improvement-2026.md
ENPIRE's broader significance is the abstraction of an agent-operable environment for physical systems. Once reset, rollout, verification, and logging are callable interfaces, a coding agent can perform physical-autoresearch rather than only offline code generation.
Related pages: physical-autoresearch, agent-operable-environments, agent-loops, harness-engineering, test-time-compute-evaluations.