← Latest papers
💻 computer science

Beyond Transformers: Linear Attention Policy for Open-Vocabulary Object Goal Navigation

This paper introduces LANav, a navigation policy that replaces standard Transformer-based self-attention with a structured linear attention mechanism—specifically enhanced by Weighted State-Expansion Linear Attention (WSLA)—to achieve superior open-vocabulary object goal navigation performance, computational efficiency, and robust sim-to-real transfer compared to existing baselines.

Original authors: Jiahong Zhang, Yifan Lin, Yandong Zhang, Sijun Shen, Kexin Wang, Yuqi Pan, Hongjuan Pei, Wei Wang, Guoqi Li

Published 2026-07-22
📖 7 min read🧠 Deep dive

Original authors: Jiahong Zhang, Yifan Lin, Yandong Zhang, Sijun Shen, Kexin Wang, Yuqi Pan, Hongjuan Pei, Wei Wang, Guoqi Li

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to find a specific object, like a "vintage toaster," in a house you've never seen before. You can only see what's directly in front of you through a narrow window, and you have to remember where you've been, what you've seen, and what you're looking for to make your next move. This is the daily life of a robot navigating the world, a field of science called Embodied AI. To do this, robots use a "brain" (a policy network) that acts like a short-term memory, constantly updating its internal state based on new sights and sounds. For a long time, scientists tried two main ways to build this memory: one that compresses history into a single, shrinking summary (like an old-school notebook), and another that keeps a long list of everything seen recently and re-reads the whole list every time to find connections (like a super-organized but slow librarian). The big question was: which method helps a robot find its way best when the object it's looking for might have a weird name or be in a completely new house?

This paper introduces a new approach called LANav (Linear Attention-based Navigation) and a special upgrade called WSLA. The researchers discovered that the "super-librarian" method (using Transformers with self-attention), which was thought to be the state-of-the-art, actually hits a wall when the robot needs to remember a long journey. Surprisingly, making the robot look at a longer history didn't help the Transformer get better; in fact, it sometimes got worse. Instead, the team found that a method called Linear Attention, which updates its memory like a steady, structured stream rather than re-reading a whole list, works much better. They took this idea and supercharged it with WSLA, which splits the memory into multiple "sub-streams" and learns how much weight to give each one. In tests, this new system found objects 36.4% of the time in a challenging simulation, beating the best Transformer models by a clear margin. Even cooler, they tested it on a real robot dog in a real room, and it succeeded 82% of the time, proving that this "streaming memory" idea works not just in computers, but in the real world.

The Problem: The Robot's Memory Crisis

Imagine you are walking through a giant, unfamiliar maze looking for a "blue chair." You can only see a few steps ahead. Every time you turn a corner, you see something new, but you also forget what you saw ten seconds ago. To find the chair, your brain needs to hold onto clues: "I passed a red door," "I heard a fan," "I turned left twice."

For years, robot scientists tried to solve this with two main types of memory:

  1. The Compressor (RNNs): These are like a robot that squishes all its past memories into one tiny, dense ball. It's fast, but as the journey gets longer, the ball gets so small and messy that the robot forgets the important details.
  2. The Re-Reader (Transformers): These are like a robot that keeps a perfect, long scroll of everything it has seen. Every time it needs to decide where to go next, it re-reads the entire scroll from start to finish to find connections. This is powerful, but it's slow and gets messy if the scroll gets too long.

The researchers asked: "If we give the robot a longer scroll (a longer history), will it get better at finding things?" They tested this with the "Re-Reader" (Transformer) approach on a task called Open-Vocabulary Object Goal Navigation. This is a fancy way of saying: "Find an object I describe to you, even if I use a weird name you've never heard before, like 'a thing that holds soup' instead of 'a pot'."

The Surprise: More History Didn't Help

The team ran a series of controlled experiments. They kept everything the same—the robot's eyes, the maps, the training rules—and only changed the memory system. They expected that if they gave the Transformer robot a longer scroll to read, it would find the objects more often.

But here is the twist: It didn't work.

When they increased the length of the history the Transformer could see, the robot's performance didn't improve. In fact, in some cases, it got slightly worse. It was as if the robot was drowning in its own memories, unable to figure out which parts of the long scroll actually mattered. The "Re-Reader" method seemed to hit a ceiling. It turns out, for a robot wandering a maze, constantly re-reading the whole history isn't the best way to update its state.

The Solution: The "Streaming" Memory

The researchers then tried a different approach: Linear Attention. Instead of re-reading the whole scroll, imagine a robot that has a "streaming memory." As it walks, it updates its internal state step-by-step, like a river flowing. It doesn't look back at the whole river; it just updates the water level based on the new rain (new observation) and the current flow.

They built a system called LANav using this streaming method. The results were immediate and impressive.

  • Better than the old ways: LANav beat both the "Compressor" (RNN) and the "Re-Reader" (Transformer) models.
  • Longer is better: Unlike the Transformer, when they gave the LANav robot a longer history to learn from, it actually got better. The longer the training history, the smarter the robot became.

The Upgrade: WSLA (The Multi-Stream Brain)

The researchers didn't stop there. They realized that even a streaming memory could be improved. They asked, "What if our robot didn't just have one stream of memory, but several, and could learn which stream to listen to?"

They created WSLA (Weighted State-Expansion Linear Attention).

  • The Analogy: Imagine the robot's brain has four different "notebooks" (sub-states) instead of one. One notebook tracks colors, another tracks shapes, another tracks turns, and another tracks sounds.
  • The Magic: A special "conductor" (learnable weighting) decides how much to listen to each notebook at any given moment. If the robot is looking for a "red box," the conductor might crank up the volume on the "color" notebook and turn down the "sound" notebook.

This WSLA system turned out to be the champion.

  • In the HM3D-OVON simulation (a massive, realistic 3D house dataset), the WSLA robot achieved a 36.4% success rate.
  • The best Transformer model only got 30.1%.
  • That's a 6.3 percentage point improvement, which is huge in this field.

Why It Matters: Distance and Real Life

The researchers also looked at how the robot navigated. They found that the new system was especially good at long journeys.

  • Short trips: Both robots did okay.
  • Long trips: The Transformer robot often got lost or gave up early. The WSLA robot, however, kept its cool. It successfully navigated distances of 15 meters or more with a success rate of 14.36%, while the Transformer only managed 6.68%.

But the real test was taking it out of the computer and into the real world. The team put their LANav system on a Unitree Go2 robot (a real, physical robot dog). They asked it to find things like a trash bin, a plant, or a chair in a real room.

  • They ran 50 trials.
  • The robot succeeded 41 times.
  • That's an 82% success rate in the real world!

This proves that the "streaming memory" idea isn't just a cool computer trick; it actually works for robots moving through real spaces, handling the messiness of real life.

The Takeaway

The paper suggests that for robots trying to find their way in complex, unknown environments, the old "re-read everything" method (Transformers) might not be the best tool. Instead, a method that updates its memory in a structured, streaming way (Linear Attention), especially when enhanced with multiple specialized sub-memories (WSLA), is far more effective. It's faster, scales better with longer journeys, and, most importantly, it actually works when you put it on a robot dog in a real room. The future of robot navigation might not be about remembering more history, but about remembering it better.

Drowning in papers in your field?

Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.

Try Digest →