Two New Research Systems Want Robots to Actually Remember Where They Are

And what's RAVEN doing differently?

The second paper, also on arXiv, proposes something called RAVEN. Same broad problem, different architecture. RAVEN stores visual embeddings directly in a vector database, tagged with pose and timestamp, and grounds retrieval to a spatial map. The key design choice is that it skips the step of converting images to text descriptions entirely.

That's worth pausing on. A lot of vision-language memory systems work by captioning everything they see and then searching through those captions later. The RAVEN team argues that's lossy. You're throwing away visual detail every time you translate an image into words. Their system keeps the raw embeddings and queries them directly.

They tested it on a Unitree Go1 robot doing long-horizon navigation in real indoor environments, which is more than a lot of papers bother with. The benchmarks show RAVEN matching frontier vision-language models on long-horizon tasks at about one-tenth the retrieval cost. That's a number worth paying attention to if you're thinking about deployment at scale.

Do these solve the actual warehouse problem?

Here's where I'd pump the brakes a little. Both systems are demonstrated in indoor navigation and question-answering tasks, which are real problems, but they're not exactly a busy fulfilment centre with forklifts, dynamic pick locations, and thirty SKUs moving every hour. The environments in these papers are complex, sure, but they're also sort of... politely complex. It's too early to say how either system would hold up when the environment is actively changing around the robot.

I called my old colleague Marcus who's been doing integrations for a large 3PL operator in the Ruhr for the past eight years, and his first question was about update latency. How fast does the scene graph refresh when something moves? Neither paper gives me a great answer on that, and that's a real limitation of what I can assess here.

There's also the question of compute. RAVEN's retrieval cost advantage is genuinely impressive on paper, but we don't know yet what the full system costs to run at the edge on something that isn't a research platform.

So should you care?

If you're in industrial automation, yes, but not because you're about to deploy either of these tomorrow. You should care because the underlying ideas, structured hierarchical memory for ObsGraph and embedding-native retrieval for RAVEN, represent the direction that robot spatial awareness is heading. The days of pre-mapped, fixed-layout automation aren't going away, but the pressure to handle dynamic and unfamiliar environments is only going up.

I've watched a lot of promising lab results not survive contact with a real factory floor. These two might be different. The architectures are sensible, the evaluation is more rigorous than average, and at least one of them has been on a real robot in a real building. That's more than I can say for a lot of what crosses my desk.

Worth watching. Not worth betting the warehouse on yet.

James Chen · 6 hours ago · 6 min