This issue: 26 selected works (25 first submissions in the window and 1 revised prior paper) · Tracks: VLN 9 / embodied agentic navigation 3 · Other directions 14 · WeChat 0
1. Key points
- There are two ways to adjust the inference budget. VLN-Flash shortens each action-decoding pass, while CHAC reruns the model less often. The latter’s success-rate drop on R2R-CE shows why the execution horizon needs calibration for each task.
- Online adaptation can reuse failed episodes. R2A uses earlier feedback for later policy updates. Read the results in light of the initial policy, feedback budget, and episode order.
- Verification is triggered by different failure signals. LeaP-VLN checks whether a subtask stop is justified, while AgenticJevNav requests review after blockage or insufficient progress. They intervene at different points in execution.
2. Deep reads
1. VLN-Flash
A single forward pass can score several short plans in a small action space.
Paper · Figure 1 · Track: VLN · Instruction following · R2R-CE / RxR-CE (continuous environments)
Problem. Multi-step autoregressive action decoding requires sequential forward passes. Executing a full action chunk can also miss a useful opportunity to observe the scene again.
Method. Tree attention scores candidate action prefixes in one forward pass. The gap between candidate scores determines how many actions to execute before the next observation. The navigation backbone stays frozen.
Evidence. On matched inputs, action decoding was 2.75× faster for Uni-NaVid and 2.57× faster for StreamVLN than autoregressive decoding. On RxR-CE val-unseen, Uni-NaVid SR rose from 55.1% to 57.8%. On a Unitree GO2, the same StreamVLN backbone completed 52/60 instructions with VLN-Flash and 49/60 with autoregressive decoding.
Assessment. Comparing the same checkpoints helps isolate the decoder’s contribution. The gains depend on the candidate action space and plan length, and deployment measurements should account for network wait separately.
Next step. Reproduce the decoding latency with a fixed execution length, then enable adaptive execution and measure SR and SPL.
2. R2A
Use both successful and failed episodes to adapt to similar instructions.
Paper · Figure 2 · Track: VLN · Online adaptation · REVERIE / R2R (discrete), R2R-CE (continuous)
Problem. Online VLN systems often use acquired feedback only for the current episode, so a later episode with a similar instruction can repeat the same mistakes.
Method. The success–failure memory retrieves episodes by instruction similarity and adds randomly sampled experiences. Policy updates favor successful trajectories and disfavor failed ones relative to a frozen pretrained prior, limiting behavioral drift.
Evidence. On discrete REVERIE val-unseen, the paper reports DUET SR rising from 46.98% to 69.84% and SPL from 33.73% to 50.98% with online adaptation. Keeping online adaptation and removing replay memory yielded 64.49% SR.
Assessment. The gains are measured under online adaptation. Episode order, the state of the replay buffer, and feedback availability affect what information the policy can use at test time.
Next step. Shuffle episodes and compare current-episode feedback, success-only replay, and joint success–failure replay under the same feedback budget.
3. CHAC
Account for navigation success alongside the model calls saved.
Paper · Figure 1 · Track: VLN · Waypoint execution · HM3D-OVON / R2R-CE (continuous environments)
Problem. Calling a vision-language model at every execution decision is expensive. Cached waypoints can lose validity after a turn, blockage, or approach to the goal.
Method. A frozen policy predicts a sequence of waypoints. Retained visual context, completion cues, and controller feedback determine when to refresh. When the policy calls for a stop, visitation history can inform at most one corrective move without another model call.
Evidence. Model calls fell by 33.4% on HM3D-OVON and 64.7% on R2R-CE, with SR drops of 0.63 and 10.82 percentage points, respectively. Across four physical tasks with 40 trials per method, CHAC succeeded 18 times and native execution succeeded 14 times.
Assessment. The simulation benchmarks show different trade-offs. Success per call alone obscures that gap. Physical trials merit a breakdown by instruction length and goal criteria.
Next step. Plot model calls against SR across instruction lengths and isolate the contribution of the terminal correction.
4. LeaP-VLN
Require evidence before moving to the next subtask.
Paper · Figure 2 · Track: embodied agentic navigation · Long-horizon instruction following · R2R-CE / RxR-CE (continuous environments)
Problem. During a decomposed route, an executor can stop early or stray from the route. Advancing without checking can propagate a local error to later subtasks.
Method. The executor jointly learns progress reasoning and action prediction. On a stop, a verifier chooses whether to advance, continue, or return to the last verified anchor, observe the scene again, and retry the current subtask.
Evidence. On R2R-CE val-unseen, the paper reports LeaP-VLN SR of 67.5% versus CorrectNav’s 65.1%, with SPL of 59.1% versus 62.3%. RxR-CE val-unseen SR was 69.6%. An AgiBot Genie 02 demonstration showed one indoor route with three subtasks.
Assessment. Recoverable anchors help contain failures, while backtracking and retries add travel. The physical demonstration shows the interface in operation, and larger trials are needed to estimate the benefit at scale.
Next step. Reproduce premature stops and off-route deviations separately, measuring verification errors, recovery success, and extra distance traveled.
5. AgenticJevNav
Give the reviewer bounded authority and call it when execution stalls.
Paper · Figure 1 · Track: embodied agentic navigation · Indoor object search · Six simulated task families and quadruped trials
Problem. Search can stall, and an action returned by remote inference can become stale when the evidence or goal changes.
Method. Blockage, execution failure, or two consecutive low-progress segments trigger Jev to review an authorized action set. A single executor checks version, freshness, and prior consumption before acting. A local RGB reflex can veto forward motion.
Evidence. Across six simulated task families and 60 trials, SR was 73.33% versus 51.67% for the strongest agent-only control. Successful trials averaged 140.55 s and 120.40 s, respectively. A GENISOM L1 succeeded in 21/30 trials in one physical setting, including tasks with a valid not-found report within budget.
Assessment. Reviewing at anomaly points makes the trigger and its authority easier to audit. The system still incurs waiting costs, and trials in one physical setting offer limited evidence of cross-scene performance.
Next step. Reproduce blockage, low progress, and stale evidence separately, logging recovery success, elapsed time, and rejected late decisions.
3. Also worth reading
- ERA · ObjectNav. Corrects occupancy errors from predicted depth with new views and verifies the target across views before stopping.
- ROAM · Outdoor semantic navigation. Maps semantic risk to terrain costs for quadruped control using open-vocabulary segmentation.
- EvoNav · Lifelong navigation. Consolidates recurring observations and search feedback into environment-specific object co-occurrence knowledge.
- RPSI-Nav · Embodied agentic navigation. Later search feedback updates both the navigation program and the rules for diagnosing it.
- EasyNav · ROS 2. Shared navigation state and timing contracts could inform the interface between a navigation agent and its controller.
- RePlan-VLA · Manipulation. Monitoring a short-horizon policy and revising later steps could transfer to navigation recovery.
4. Trends and actions
Trends
- Different stages offer different opportunities to reduce inference cost. VLN-Flash addresses decoding cost per call, while CHAC addresses refresh frequency. Measure these separately before combining them into end-to-end time.
- Verification is becoming a testable interface. LeaP-VLN checks subtask transitions, while AgenticJevNav reviews execution anomalies. Each allows separate measurement of missed and spurious reviews.
Research gap
- A shared long-horizon continuous benchmark is needed to test when an independent verifier should intervene, how much delay it can tolerate, and when control should pass back to the local executor before the budget runs out.
Next steps
- High · Reproduce VLN-Flash and CHAC with fixed backbones and hardware, recording latency per decode, model calls, and SR/SPL separately.
- Medium · Evaluate LeaP-VLN and AgenticJevNav on the same long-horizon tasks with premature stops, blockages, and delayed feedback. Compare recovery costs at transitions and after anomalies.
- Medium · Reproduce R2A from a cold start and with existing replay memory. Clear memory for a control in new scenes.
5. Full index
The 26 selected works are grouped by the two navigation tracks (A), transferable methods (B), and adjacent fields (C).
VLN and semantic navigation · 9 papers
- VLN-Flash. See the deep read.
- R2A. See the deep read.
- CHAC. See the deep read.
- ERA. Uses new viewpoints gained through motion to correct mapping errors.
- ROAM. Maps semantic risk to terrain costs for a quadruped.
- EvoNav. Consolidates short-term observations into lasting environment knowledge.
- SearchDiffusion. Guides object search trajectories with a spatial utility map.
- ROS 2 semantic navigation. Connects open-vocabulary detection to Nav2 in Gazebo.
- Racing the Clock. Revised comparison of teachers for budget-constrained object search.
Embodied agentic navigation · 3 papers
- LeaP-VLN. See the deep read.
- AgenticJevNav. See the deep read.
- RPSI-Nav. Updates search rules and the mechanism that produces them.
Transferable methods · 8 papers
- EasyNav. Connects ROS 2 navigation modules through an explicit shared state.
- RePlan-VLA. Tracks long-horizon manipulation progress and revises the plan from feedback.
- AgentRoboCritic. Attributes action failures against local intent.
- CoachSI. Uses a language coach to obtain successful rollouts for relabeling.
- CRA. Integrates continuous semantic risk into a visual-navigation safety filter.
- FedDIA. Weights federated updates by environment difficulty and improvement.
- Hierarchical Skill Composition. Reuses pretrained skills in unseen FourRooms layouts.
- Homeward. Reprojects changed camera observations into the training view.
Adjacent fields · 6 papers
- Aerial navigation (3) AeroRSI, Fly as Instructed, CityNav-VLA
- VLA manipulation (2) A²-VLA, ScaleVLA
- Autonomous driving (1) Ego2Flow