AuthorsPeiqi Liu, Zhanqiu Guo, Mohit Warke et al.
TL;DR
DynaMem uses an online dynamic 3D voxel memory with add-remove updates plus hybrid VLM and mLLM querying to reach 70% pick-and-drop success on non-stationary objects vs 30% for OK-Robot.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing open vocabulary mobile manipulation systems assume static environments, leading to only 30% success on non-stationary objects compared to 70% with DynaMem.
When objects move, static spatio semantic memory fails to localize goals, so navigation and manipulation stacks like OK Robot frequently chase outdated locations.
HOW IT WORKS
DynaMem combines a Dynamic 3D Voxel Map, Embedded Vision Language Features, Multimodal Large Language Models, and Exploration Primitives to maintain and query online spatio semantic memory.
You can think of DynaMem like RAM blocks that constantly rewrite 3D feature cells, while an mLLM acts as a librarian pointing to the most relevant recent snapshots.
This KEY_MECHANISM of add remove voxel updates plus hybrid VLM mLLM grounding lets DynaMem forget outdated geometry and abstain on missing objects, which a plain context window cannot.
DIAGRAM
This diagram shows how DynaMem processes a text query, selects images via voxel similarity or mLLM QA, and confirms presence with OWL v2 before returning 3D coordinates.
DIAGRAM
This diagram shows how DynaMem is evaluated on DynaBench with different query variants and ablations against human performance.
PROCESS
Dynamic 3D Voxel Map
DynaMem converts posed RGB D images into a sparse voxel grid, storing location, observation count, image ID, semantic features, and latest observation time.
Adding Points
DynaMem backprojects depth, segments with SAM v2, aggregates CLIP or SigLIP features, and updates voxel counts and averaged features for newly observed points.
Removing Points
DynaMem ray casts through the camera frustum, projecting voxels and deleting those that should be occluded but are not, handling moved or removed objects.
Querying DynaMem for Object Localization
DynaMem uses Embedded Vision Language Features or Multimodal Large Language Models plus OWL v2 to find the latest valid image and return a 3D object location or abstain.
KEY CONTRIBUTIONS
Dynamic 3D Voxel Memory for changing environments
DynaMem introduces a Dynamic 3D Voxel Map with add and remove operations, storing observation counts, image IDs, semantic features, and timestamps for online updates.
Hybrid VLM and mLLM querying with abstention
DynaMem combines Embedded Vision Language Features, Multimodal Large Language Models, and OWL v2 cross checks to localize objects or explicitly return object not found.
DynaBench dynamic benchmark and real robot evaluation
DynaMem is evaluated on DynaBench with nine environments and on Stretch SE3, achieving 70% success on non stationary objects versus 30% for OK Robot.
RESULTS
Pick and drop success
70% success
+40 percentage points over OK-Robot static voxelmap baseline at 30% success
Dynamic object navigation failure
6.7% failure
-46.6 percentage points vs OK-Robot dynamic object navigation failure at 53.3%
Human success on DynaBench
81.9% success
DynaMem hybrid querying reaches 74.5%, 7.4 percentage points below humans
Hybrid query success
74.5% success
+3.9 percentage points over VLM feature default at 70.6% on DynaBench
These numbers come from real Stretch SE3 experiments and the DynaBench offline benchmark, which test dynamic 3D visual grounding and open vocabulary mobile manipulation. MAIN_RESULT shows that DynaMem makes dynamic environments tractable, closing much of the gap to human performance while doubling success over static OK-Robot.
BENCHMARK
These numbers come from real Stretch SE3 experiments and the DynaBench offline benchmark, which test dynamic 3D visual grounding and open vocabulary mobile manipulation. MAIN_RESULT shows that DynaMem makes dynamic environments tractable, closing much of the gap to human performance while doubling success over static OK-Robot.
BENCHMARK
Success rate for open vocabulary pick and drop tasks with changing object locations.
BENCHMARK
Success rate on DynaBench for different DynaMem query variants and human participants.
KEY INSIGHT
On DynaBench, DynaMem with simple VLM feature querying already reaches 70.6% success, only 11.3 percentage points below human performance at 81.9%.
This is surprising because one might expect complex mLLM QA alone to dominate, yet DynaMem’s voxel feature pipeline plus OWL v2 cross check is already highly competitive.
WHY IT MATTERS
DynaMem enables robots to maintain and query a live 3D spatio semantic memory that gracefully handles objects appearing, moving, and disappearing over time.
Builders can now design open vocabulary mobile manipulation systems that explore, re search, and update maps online, instead of assuming static pre mapped scenes or blindly trusting outdated observations.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…