Adelaide team automates embodied agent architecture search

Automated architecture search has been a quietly persistent thread in language-model work, but embodied AI stacks (the perception, memory, planning, and action pipelines behind navigation robots and manipulation policies) have almost universally been hand-designed. A team from the Australian Institute for Machine Learning at the University of Adelaide argues in a new arxiv paper that this is now testable ground rather than aspirational.

They introduce two pieces. AgentCanvas is a typed-graph runtime that treats an embodied executor as an editable node-and-wire program with simulator-aware execution and episode-level logs. KDLoop is a coding-agent search procedure that cycles through proposal, critique, experiment, and distillation, with reflection triggered after stalls. The pair are evaluated across three AAS variants and four embodied executors spanning vision-language navigation (MapGPT and SmartWay on Matterport3D and Habitat), embodied question answering (ExploreEQA on HM3D), and language-conditioned manipulation (VoxPoser on LIBERO).

The gains are directional rather than transformative. MapGPT reportedly moved from a 46.9±3.1% baseline to approximately 54% under search-driven edits, and ExploreEQA gained +4.7% under an AFlow variant and +3.0% under KDLoop over a 43.0±1.7% baseline. Take the specifics as reported, not settled, since these are simulator numbers on published benchmarks and the paper itself describes the improvements as deployable and directional rather than a clean win.

What the paper is honest about is why this is hard. Optimization signals get masked by rollout noise, search gets trapped in what the authors call local edit basins, and episode-level credit assignment only partially emerges even when the logs are detailed. Those aren't incidental engineering problems, they are the reasons hand-designed stacks have held on. What the reporting doesn't give you is compute cost per search run, a wall-clock comparison against a human engineer iterating manually, or any out-of-simulator evaluation, and any of those would change the picture.

The forward-looking part is who this helps. If AAS becomes a real tool for embodied agents, small robotics and simulation groups without a dedicated architecture team could inherit a lot of the tuning that today lives inside senior researchers' heads. The tell over the next year will be whether anyone reproduces the gains outside Matterport3D and LIBERO, and whether the KDLoop reflection loop generalises to real-robot logs rather than simulator ones.