REA Runtime Splits Robot Agents Into Brain and Cerebellum

Researchers from the National University of Singapore, UC Berkeley, and the University of Tokyo have officially released REA, a three-level runtime architecture that achieves 95.9% success on the LIBERO-Pro benchmark while reducing control steps by 25% and control time by 17%. The system decouples slow deliberative foundation models from fast reactive policies, ensuring that no running policy ever waits on the brain.

Architecture and Design Philosophy

REA introduces a radical split: foundation models serve as the deliberative "brain," while vision-language-action models and whole-body controllers act as the "cerebellum." Between them sits an embodied router that manages policy selection without blocking execution. The core innovation is a precomputed route memory that stores each policy's measured success offline, eliminating the need for expensive foundation model calls during inference.

The runtime employs a boundary-triggered judge gate that consults the brain only between policy executions, never during them. This design cuts 2.6 seconds per episode compared to blocking advisors, roughly 20% of total control time. A shared-budget escalation ladder provides bounded recovery when policies fail, maintaining system responsiveness even under challenging conditions.

Benchmark Performance and Real Hardware Results

Across 3,000 test episodes spanning 60 LIBERO-Pro tasks, REA outperformed the strongest single-agent baseline, which achieved 94.7% success. The system demonstrated consistent gains in efficiency metrics while maintaining higher success rates. On the RoboTwin 2.0 benchmark across 100 episodes, REA with VLA policies hit 84% success, rising to 86% when combined with code-generation capabilities, surpassing the MEDex baseline at 80%.

Real-world validation proved equally compelling. On a Unitree G1 humanoid walk-to-push task, REA scored 8 of 10 trials versus 5 of 10 for Agents SDK and 2 of 10 for ASPIRE, with statistical significance at p=0.023. The system achieved perfect 10 of 10 success on an AlohaMini2 block placement task, demonstrating reliability across different robotic platforms and manipulation challenges.

Development Costs and Efficiency Gains

The route memory approach delivers substantial efficiency advantages during development. According to the authors' cost analysis, building the route memory required 1,790 episodes with zero token consumption. In contrast, a Codex plus GPT-6 approach consumed 89.0 million input tokens, highlighting the computational savings of offline policy success measurement.

The non-blocking design means inference runs at the speed of the fastest applicable policy rather than waiting for foundation model deliberation. This architectural choice translates directly to faster episode completion times and more responsive robot behavior, critical factors for real-world deployment where latency impacts safety and task success.

Release Date and Availability

The REA runtime was officially released on November 2, 2026, with the research paper published on Hugging Face. The release includes implementation details for the three-level architecture, route memory construction methods, and benchmark evaluation protocols. The system transfers across multiple embodied AI benchmarks including BEHAVIOR-1K, LIBERO-Pro, and RoboTwin 2.0.

What This Means

REA represents a significant step toward practical embodied AI systems that balance deliberative reasoning with reactive control. By eliminating the waiting penalty that has plagued brain-policy architectures, the runtime makes foundation model guidance feasible for time-sensitive robotic tasks. The strong real-hardware results on both humanoid and manipulation platforms suggest the approach generalizes beyond simulation, a persistent challenge in embodied AI research. For teams building production robot systems, REA offers a proven architecture that maintains high success rates while reducing both inference latency and computational costs.