Global Prior Meets Local Consistency: Dual-Memory Augmented Vision-Language-Action Model for Efficient Robotic Manipulation
New AI model for robots uses two memory systems to achieve 2.9x faster inference and major performance gains.
A research team led by Zaijing Li has introduced OptimusVLA, a novel dual-memory architecture designed to overcome critical bottlenecks in robotic manipulation AI. Current Vision-Language-Action (VLA) models suffer from low inference efficiency due to a distribution gap between noise priors and target actions, and poor robustness from ignoring historical context. OptimusVLA addresses these with two specialized memory systems: a Global Prior Memory that retrieves task-level priors from similar past trajectories to shorten the generative path, and a Local Consistency Memory that models executed action sequences to enforce temporal coherence and smoothness.
The technical breakthrough shows remarkable results across multiple benchmarks. In simulation, OptimusVLA achieved a 98.6% average success rate on LIBERO, improved over the baseline π₀ by 13.5% on CALVIN, and attained 38% success on the challenging RoboTwin 2.0 Hard. Real-world evaluations were even more impressive, with the model ranking best on Generalization and Long-horizon tasks, surpassing π₀ by 42.9% and 52.4% respectively while delivering a 2.9x inference speedup. This represents a significant step toward more efficient, reliable, and context-aware robotic systems capable of complex manipulation tasks.
- Achieved 98.6% success rate on LIBERO benchmark and 2.9x faster inference in real-world tests
- Introduced dual-memory system: Global Prior Memory for task priors and Local Consistency Memory for temporal coherence
- Surpassed baseline π₀ by 42.9% on generalization tasks and 52.4% on long-horizon tasks in real-world evaluation
Why It Matters
Enables more reliable and efficient robots for manufacturing, logistics, and home assistance by dramatically improving success rates and speed.