ENPIRE

ENPIRE (Environment, Policy Improvement, Rollout, Evolution) is a harness framework for coding agents that turns real-world robot policy learning into a controllable optimization procedure that agents can manage. It was developed at NVIDIA GEAR Lab (Generalist Embodied Agent Research) in collaboration with Carnegie Mellon University and UC Berkeley. It was announced on June 17, 2026 during GTC 2026. NVIDIA has indicated plans to release it as open source.

The four system modules

• Environment (EN) — automatic reset, verification, safety, and logging interfaces that the agent can call.
• Policy Improvement (PI) — generates and revises policy code from rewards, videos, traces, and failure cases. Supports multiple regimes: heuristic learning, tool calling, behavior cloning, offline and online reinforcement learning.
• Rollout (R) — runs budgeted physical trials on a single robot or in parallel across a fleet, preserving state, action, video and result for later audit.
• Evolution (E) — compares branches of ideas, reuses successful recipes, and prunes hypotheses that fail on hardware. Agents analyze logs, consult the literature, and improve training infrastructure and algorithm code.

Results

Powered by ENPIRE, frontier coding agents autonomously developed policies reaching a 99% pass@8 success rate on challenging real-world dexterous manipulation tasks: Push T, organizing pins into a pin box (Pin Insertion), GPU insertion into a motherboard socket, and tying and cutting a zip-tie. Experiments ran on a fleet of eight dual-arm robots. The authors propose two efficiency metrics for multi-agent physical autoresearch: Mean Robot Utilization (MRU) and Mean Token Utilization (MTU).

Agent evaluation

The team evaluated three coding agents on the AutoEnvBench benchmark: Codex (GPT-5.5), Claude Code (Opus 4.7), and Kimi Code (Kimi K2.6). The benchmark tracks agent-driven research progress over wall-clock time on Push-T (heuristic learning) and Pin Insertion (gradient-based learning) tasks.