Relevant memory route
The recalled avoid_region intent blocks the prior hazard.
SUCCESSAgentic memory for Physical AI
Persistent memory that changes the next action.
CoreSense Demo uses Amazon Bedrock to embed a robot's prior failure and CockroachDB Cloud to persist and retrieve it after a process restart. The recalled recommendation changes a deterministic route from failure to success.
Persistent memory timeline
Retrieval suggests a bounded intent. A deterministic executor decides how that intent changes the public demonstration route.
01Observe
Load the synthetic situation.
02Fail
Encounter unsafe terrain.
03Persist
Write memory and vector.
04Restart
End the first process.
05Recall
Run semantic retrieval.
06Apply
Use avoid_region.
07Succeed
Take the safer route.
Memory changes behavior
Click each condition to inspect the route and recalled memory state.
Recalled failure memory
- Memory ID
- memory-demo-winter-grid-v1
- Failure mode
- unsafe_traversability
- Recommendation
- avoid_region
- Evidence ID
- demo-evidence-winter-grid-v1
- Reference policy
- Applied
This is an opaque demo evidence identifier, not a production integrity proof. The public repository does not implement or claim a private production integrity-verification protocol.
Applications, examples & public references
A visual explanation of where persistent failure memory can support field robotics. External media is clearly separated from this repository's synthetic development fixture.
7 × 5 SYNTHETIC GRID
Included fixture
Failure → persistence → recall → replan
The only scenario executed by this public repository.
Public robotics reference ↗
TartanGround / ANYmal
A relevant public source for future terrain-memory evaluation.
Official overview video ↗
Multimodal ground robot data
External public media; no video is redistributed here.
BEDROCK → COCKROACHDB
CoreSense cloud path
Semantic memory across process restarts
Embedding, persistence, vector recall, and bounded replanning.
No memory → failure
Relevant memory → successful replan
Process restart → memory retained
Unrelated query → no forced action
Industry / Customer Use Case Illustrative — not a deployment claim
Illustrative field-robot inspection workflow
A legged robot inspecting a winter forest, industrial site, or remote facility may encounter an unsafe patch of terrain. A conventional stateless run can repeat the same failure after a software restart. CoreSense demonstrates how the failure can be stored as a searchable memory and recalled when a similar situation appears again.
What changes after recall
- The stored memory recommends the bounded intent
avoid_region. - The public reference executor adds the prior hazard cell to its blocked set.
- The deterministic planner selects a different route.
- The second run reaches the goal without revisiting the known hazard.
The repository does not claim that ANYbotics, ETH Zurich, TartanGround, or a named customer uses CoreSense. TartanGround is an external public research reference. The included executable scenario remains synthetic.
Model and data context
CoreSense is a memory infrastructure demo, not a locomotion-model training result. A future evaluation can use public robot trajectories or learned terrain features as memory inputs while keeping policy training and memory retrieval as separate layers.