HTCE gives navigation AI a simpler way to reuse past routes

Remembering the route, not the floor plan

An agent searching for a bed in an unfamiliar home cannot simply replay movements from another building. Rooms and doorways occupy different positions. But a successful search may contain a reusable pattern: leave a living room, follow a passage, and enter a bedroom.

In an unreviewed preprint, Xiaoran Xu and colleagues propose HTCE, a framework for organizing and retrieving these patterns. Their central idea is to make differently described navigation experiences comparable before using them to guide an agent. The reported results come from simulated indoor navigation, not physical robot trials.

Turning descriptions into comparable patterns

HTCE stores experience as a graph: regions form the nodes, and observed connections form the edges. Records also contain object sightings and successful approaches to targets.

The challenge is that similar approaches can produce different records. One passage might be labeled “hallway,” while another appears as “entryway” followed by “corridor.” Repeated observations can add intermediate nodes without introducing a meaningful new choice.

HTCE addresses this through two reductions. First, it maps compatible labels to shared semantic roles and contracts consecutive repetitions in route descriptions. In the paper’s configuration, corridor, hallway, entryway, and unknown room share a transition role. Rooms such as bedrooms and kitchens remain semantic anchors.

Second, it simplifies the graph by suppressing intermediate transition nodes that have exactly two connections and are not route endpoints. Junctions remain because they represent choices. Semantic anchors also remain, even when they have only two connections.

Importantly, a direct room-to-room connection stays distinguishable from one containing a passage. Simplification also retains access to the original observations.

Memory guides a frozen model

The framework borrows its organizing principle from graph homeomorphism: adding intermediate points along a connection need not change the underlying connectivity. HTCE adds semantic restrictions because, for navigation, not every intermediate region is disposable.

During a search, the system compares current candidate approaches with stored successful ones. Retrieved experience supplies semantic guidance, while current color-and-depth observations and visible action candidates determine executable movements.

The vision-language model remains frozen; its parameters are not updated during navigation. Learning from experience here means updating explicit memory, rather than retraining the model. HTCE also distinguishes objects seen locally from objects seen through an opening, preserving information about where a target may actually be located.

What the experiments report

The authors evaluated HTCE in Habitat ObjectNav using three indoor benchmark settings. All compared systems used the same frozen Qwen3-VL-Plus backbone, action proposer, controller, and 40-step interaction budget. External baselines were reproduced and adapted by the researchers, so these results describe that controlled implementation rather than an unrestricted comparison with the original systems.

They report these success-rate and success-weighted path-length scores, respectively:

• HM3D v0.1: 59.9% and 31.8%, across 2,000 episodes.
• HM3D v0.2: 71.0% and 31.5%, across 1,000 episodes.
• MP3D: 46.6% and 16.6%, across 2,195 episodes.

Success rate measures target completion. Success weighted by path length, or SPL, also rewards efficient paths. HTCE had the highest reported values on both measures in each setting, although some margins were small: its HM3D v0.1 SPL exceeded WMNav by just 0.1 percentage point.

Removing successful-approach matching reduced SPL by 2.8 points on HM3D v0.1 and 1.3 points on v0.2. Those component comparisons support the authors’ argument that retrieved route structure contributes to efficiency, within their experimental setup.

What “lossless” does—and does not—mean

Across 927 recorded successful routes from two episode orders, canonicalization reduced 130 and 140 distinct anchor-bounded descriptions to 107 patterns each. That represents index reductions of 17.7% and 23.6%, not equivalent reductions in total memory usage.

The authors describe this compression as lossless because their defined reductions preserve specified navigation structure and retain underlying evidence. Controlled checks supported those prescribed relationships, and the full system’s navigation scores were not lower than the corresponding no-reduction variant.

However, the exact structural definition is richer than the descriptor used for fast online retrieval. The paper explicitly shows cases where that descriptor cannot distinguish swapped branch positions or different numbers of same-role branches, although exact graph comparison can.

The semantic label partition also determines which distinctions are absorbed. Together with the simulated evaluation and adapted baselines, these boundaries limit the conclusion: HTCE offers evidence for useful, compact experience reuse under the tested conditions, not proof that every navigation-relevant difference survives practical retrieval.