AuthorsJingwei Zhang, Lei Tai, Ming Liu et al.
TL;DR
Neural SLAM embeds SLAM-like motion prediction and measurement update into an external memory, enabling 13.732 average reward and 46/50 success on 16×16 exploration tasks.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Neural SLAM targets exploration where agents must clear all accessible cells before 750 steps, but random policies need 5531.600 steps on 16×16 worlds.
Without long term memory and cognitive mapping, reinforcement learning agents miss distant unexplored regions, leaving coverage tasks unsolved and wasting many actions.
HOW IT WORKS
Neural SLAM uses an LSTM, Localization and Motion Prediction, Data Association, Measurement Update, and Mapping over a 2D external memory to maintain a global map-like belief.
You can think of Neural SLAM as a learned SLAM stack where the LSTM is the controller and the external memory is a writable grid RAM that stores the agent’s cognitive map.
This embedded SLAM structure lets Neural SLAM plan using a persistent internal map instead of a short context window, enabling coverage of large, partially observed environments.
DIAGRAM
This diagram shows how Neural SLAM updates its external memory each time step using motion prediction, data association, and measurement update before computing a policy.
DIAGRAM
This diagram shows how Neural SLAM is trained with A3C on grid worlds and then evaluated on larger unseen environments.
PROCESS
Localization and Motion Prediction
Neural SLAM uses Localization and Motion Prediction to apply the previous action onto the access weights, updating the belief over positions on the global map.
Data Association
Neural SLAM computes a key from the LSTM and runs Data Association against the external memory using cosine similarity to obtain content based access weights.
Measurement Update
Neural SLAM interpolates motion and content weights, applies a shift kernel, and sharpens them during the Measurement Update to refine the current belief.
Mapping and Policy
Neural SLAM performs Mapping with erase and add vectors, reads a summary vector, and feeds it with the LSTM state into the policy and value heads.
KEY CONTRIBUTIONS
Neural SLAM architecture
Neural SLAM embeds Localization and Motion Prediction, Data Association, Measurement Update, and Mapping into an external memory controller to evolve SLAM like behaviors end to end.
Long term exploration memory
Neural SLAM uses a 2D external memory of size 16×16×16 as a cognitive map, enabling coverage of grid worlds up to 16×16 under a 750 step limit.
Generalization to larger worlds
Neural SLAM trained on 8×8 to 12×12 worlds generalizes to 16×16, achieving 13.732 average reward and 46/50 success episodes compared to 7.196 and 37/50 for A3C Nav2.
RESULTS
Average reward
13.732
+6.536 over A3C-Nav2
Average steps
174.920
−108.560 steps vs A3C-Nav2
Success ratio
46/50
+9 episodes vs A3C-Nav2
Random steps
5531.600
baseline difficulty for random exploration
On 50 randomly generated 16×16 grid worlds, Neural SLAM is evaluated for exploration coverage under a 750 step cap, demonstrating substantially higher reward and success than A3C-Nav2 and A3C baselines.
BENCHMARK
On 50 randomly generated 16×16 grid worlds, Neural SLAM is evaluated for exploration coverage under a 750 step cap, demonstrating substantially higher reward and success than A3C-Nav2 and A3C baselines.
BENCHMARK
Average reward on 50 randomly generated 16×16 environments.
KEY INSIGHT
Neural SLAM with an explicit motion model reaches 13.732 reward, while A3C-Ext with an external memory but no motion model drops to −8.127.
This is surprising because both agents have similar memory capacity, yet without embedded SLAM structure the external memory actually harms performance on larger worlds.
WHY IT MATTERS
Neural SLAM shows that embedding SLAM like motion prediction and measurement update into external memory yields robust long horizon exploration policies.
Builders can now design agents that learn their own cognitive maps for coverage and navigation tasks, without hand engineered SLAM pipelines or explicit occupancy grid maps.
Guilin Zhang, Wei Jiang et al.
· 2026
A-MAC scores candidate memories using Utility, Confidence, Novelty, Recency, and Type Prior combined by a learned linear admission policy with Algorithm 1 A-MAC Memory Admission. On the LoCoMo benchmark, A-MAC achieves F1 0.583 and 2644 ms latency, improving F1 by 0.042 and reducing latency by 1187 ms compared to A-mem.
Robbyant Team, Zelin Gao et al.
arXiv 2026 · 2026
LingBot-World combines a Data Engine, Fundamental World Model, Action-Conditioned World Model, and Post-Training causal adaptation to turn a 28B-parameter video generator into a real-time interactive world simulator. On the VBench benchmark, LingBot-World achieves a dynamic degree of 0.8857 versus 0.7612 for Yume-1.5, while also improving imaging quality to 0.6683.
Manoj Madushanka Perera, Adnan Mahmood et al.
· 2026
AgenticAI-DialogGen chains ChatPreprocessor, KnowledgeExtractor, TopicAnalyzer, KnowledgeGraphBuilder, PersonaGenerator, DuelingChat Agent, ConversationValidator, ConversationRefiner, QAGeneration, and PostProcessing to turn raw multi-session chats into topic-guided, persona-grounded conversations with explicit short- and long-term memories. On the TGC / KG memory QA benchmark, Mistral-7B fine-tuned within AgenticAI-DialogGen achieves 87.36 F1, compared to GPT-4’s 83.77 F1 in a zero-shot setting on the same task.
Answers use this explainer on Memory Papers.
Checking…