A navigation system that adapts while it moves recorded higher scores than a cited baseline across several unseen-environment tests, according to a version-one arXiv preprint. The proposed Credit-Guided Policy Improvement method leaves the pretrained navigation backbone frozen, uses the observation produced by each action to guide a lightweight update, and keeps that update only after a verification step. The benchmark results are reported as point estimates, and the paper's real-world evidence consists of qualitative robot trials.
A system that judges its own updates
The question is whether feedback-free test-time adaptation can use action-induced observation transitions and selective update verification to improve vision-language navigation in unseen environments. In the paper's setup, the system receives no external task-level correctness or outcome signal, expert actions, target locations, dense rewards, trajectory-level success labels or human feedback. After a movement action, its new environmental information is the resulting observation.
CGPI, the paper's name for the method, derives signed, reference-relative decision credit from the observation that follows the executed action. It proposes a lightweight reranker update while leaving the pretrained backbone untouched, then tests the update before retaining it or rolling it back. The design is intended for adaptation when an agent can observe what happened next but cannot be told whether the overall task is correct.
The experiments covered the REVERIE, R2R and R2R-CE vision-language navigation benchmarks, using the HAMT, DUET, BEVBert and ETPNav backbones under standard splits and continual test-time adaptation protocols.
Higher scores across three benchmarks
On REVERIE, CGPI had the best reported performance across the Val Seen, Val Unseen and Test Unseen splits for both HAMT and DUET. With DUET on Val Unseen, its SR/SPL scores, the paper's navigation-success and path-efficiency measures, were 58.76/40.15, compared with 56.92/38.03 for IDEA. On DUET Test Unseen, the corresponding figures were 57.31/41.87 versus 55.12/39.84; on HAMT Val Unseen, they were 36.97/33.26 versus 34.92/31.52.
CGPI also posted higher reported SR/SPL pairs in the cited Val Unseen comparisons for R2R and R2R-CE. On R2R, the DUET comparison was 79/68 versus 76/67, and the BEVBert comparison was 79/70 versus 76/68. On R2R-CE, BEVBert was 65/54 versus 62/52, while ETPNav was 61/52 versus 59/51. The paper reports no confidence intervals, inferential tests or repeated-run variability, so these remain point estimates rather than a formal estimate of uncertainty.
The safeguards were tested too
To examine the components, the authors compared variants with and without grounding and reference-relative credit. On REVERIE Val Unseen with DUET, the combined variant recorded SR 58.76, SPL 40.15, Credit-Progress Correlation of 0.572 and Wrong Update Rate of 0.154, the strongest reported combination across the four variants. CPC and WUR were post-hoc diagnostics, however, and were not available to the system during adaptation.
Verification and the candidate-level KL constraint were also tested separately. With full CGPI on the same REVERIE and DUET split, SR/SPL was 58.76/40.15, policy drift was 0.043, the accepted improvement rate was 0.741 and the retention rate was 61.3%. Without verification, the figures were 57.41/38.96, 0.071, 0.548 and 100.0%, respectively. Removing candidate KL control produced 58.29/39.72, policy drift of 0.057 and an accepted improvement rate of 0.704, with retention at 67.8%.
A separate post-hoc counterfactual analysis compared CGPI with direct commitment. It reported an accepted improvement rate of 0.741, an accepted improvement gain of 0.63 metres and a rejection precision of 0.716 for CGPI, against 0.548 and 0.19 metres for direct commitment. The paper says that 71.6% of rejected candidates would have produced negative short-horizon oracle progress. Those diagnostics used geodesic information that was unavailable during adaptation, so they assess candidate selection after the fact rather than describe feedback the agent received in operation.
Continual adaptation also had higher reported scores than resetting the method after each episode, and both versions were above the frozen HAMT result in one comparison. On REVERIE Test Unseen, frozen HAMT recorded SR/SPL 30.40/26.67, episode-reset CGPI recorded 32.62/28.41, and continual CGPI recorded 34.73/30.78. In this comparison, the continual version had the highest reported pair, although the paper does not report uncertainty intervals for it.
A robot demonstration, not a field result
The paper ends with three qualitative Unitree Go2 trajectories. They are reported as successful multi-stage instruction-following trials and are described as illustrating sim-to-real feasibility without real-world fine-tuning or external outcome feedback. Because this evaluation is qualitative, it does not establish a quantitative real-world success rate or show robustness beyond the named robot demonstrations.
The work is an arXiv preprint, version 1, dated 29 September 2026, with no journal or repository venue reported in the supplied metadata.
Paper data and sources
Original title: Credit-Guided Policy Improvement for Test-time Adaptive Vision-Language Navigation
Authors: Yang Li, Sijia Zhang, Yihan Li et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-09-29
DOI: Not available
Original paper · Full text