ABot-AgentOS: Universal Robot Runtime Layer

acvlab · hf · 2026-07-14

ABot-AgentOS: Universal Robotic Agent OS

The authors propose a universal robot runtime layer sitting above low-level controllers, designed for long-horizon embodied tasks involving:

Additionally, they introduce EmbodiedWorldBench, an executable benchmark featuring 16 indoor, outdoor, and mixed scenes, 4 difficulty levels, and 200+ tasks covering navigation, object finding, NPC dialogue, dynamic events, and trace-grounded scoring.

Memory & Self-Evolution

The system also proposes Universal Multi-modal Graph Memory: transforming conversations, visual observations, spatial contexts, temporal relations, and task trajectories into typed nodes and edges to serve as a persistent and auditable memory foundation. There is also a "failure-driven" self-evolution loop that converts diagnosed memory failures into gated runtime evo-assets, enabled only in subsequent evaluation splits to avoid ground-truth leakage in the current split.

Experimental Results

On an initial subset of EmbodiedWorldBench, ABot-AgentOS improved task success rate and goal completion compared to a single-controller baseline. On several memory benchmarks, the static version achieved:

Self-evolution further boosted LoCoMo to 88.7, OpenEQA to 60.4, and Mem-Gallery to 89.0.

Original post →

More from Embodied

Embodied channel →