Keep pulling the thread on Jim Fan.
The ASPIRE system surpasses prior methods by up to 77% on the LIBERO-Pro manipulation benchmark under perturbation.
The ASPIRE system surpasses prior methods by up to 72% on the Robosuite bimanual handover benchmark.
The ASPIRE system surpasses prior methods by up to 32% on the BEHAVIOR-1K long-horizon household tasks benchmark.
On the LIBERO-Pro Long benchmark, the ASPIRE system achieves a 31% success rate in zero-shot generalization to unseen long-horizon tasks, compared to 4% for prior methods.
Skills discovered by the ASPIRE system in simulation can be transferred to real-world robots, reducing programming effort across different embodiments and robot APIs.
ASPIRE is a continual learning system that autonomously writes and refines robot control programs using a code-as-policy paradigm and compounds experience into a reusable skill library.
The ASPIRE system discovers skills that persist across different tasks, simulation and real-world settings, and robot embodiments.
The ASPIRE system includes a closed-loop robot execution engine that uses fine-grained multimodal traces for autonomous failure diagnosis, repair synthesis, and validation.
The ASPIRE system features a continually expanding skill library that distills validated fixes into reusable and transferable knowledge.
The ASPIRE system uses evolutionary search to generate diverse task sequences and control programs to explore beyond single-trajectory refinement.