The same request can call
for a different plan.
Memory must persist and remain revisable.
A harness around a frozen VLA
CarryOn.
A Dual-Loop Harness for Personalized and Reliable Robot Assistance
Remember the person and the scene. Check actions before contact. Recover from failures and retain what works.
Memory-assisted scene selection88/88
Correct dispatches with either text or image references.
RoboLab · planning context ↗Violation rate52.6→6.1%
Runtime shielding reduces violations beyond safe demonstrations.
LIBERO-Safety · proactive shield↗Recovery with refined skills13→80%
VLA alone → two reflection rounds; cumulative over 200 episodes.
LIBERO-Object · reactive recovery↗01 / Motivation
Action capability is a beginning.
Assistance asks for more.
Everyday assistance requires choosing the right task for the user and carrying it out reliably as the world changes.
The same request can call
A valid plan can still
lead to unsafe or failed motion.
Protect before contact. Recover after failure.
02 / Method
Two loops.
One continuing task.
Planning selects the task; execution protects and repairs it. Visual feedback connects both around a frozen VLA.
What stays fixedVLA weightsWhat can improveUser / scene context & recovery skill versions
03 / Personalization
Remembered context
becomes the next decision.
Remember preferences, locate personal objects, and revise both the plan and memory when new evidence arrives.
RoboLab / four scenes
Less search.
Faster planning.
After memory becomes stale
Correct the memory.
Use it again.
| Query | Memory corrected | Follow-up dispatch |
|---|---|---|
| Text | 76 / 88 | 83 / 88 |
| Image | 77 / 88 | 85 / 88 |
Update memory by recording the new location and removing the stale association.
On the real robot
Context changes what gets dispatched.
04 / Proactive protection
Check the motion
before contact happens.
Predict gripper and carried-object motion, check obstacle distances, then release actions or retreat.
LIBERO-Safety / 900 episodes per system
See the interventions
05 / Reactive recovery & reflection
Repair the failure.
Keep the experience.
Control-based skills restore task progress. Offline reflection refines their parameters and procedures for later attempts.
LIBERO-Object / 12 cm object displacement
Cumulative task completion
100%50%0
13.0VLA only
5.0CaP-X
50.5Harness
VLA
VLA
29.0R0
70.5R1
80.0R2
R0 Shared base skill
58 / 200 solvedR1 One reflection round
83 / 142 solvedR2 Two reflection rounds
19 / 59 solved
58 / 200 solvedR1 One reflection round
83 / 142 solvedR2 Two reflection rounds
19 / 59 solved
LIBERO-Recovery / 770 recorded failure states
Recovery from recorded failures
100%50%0
32.1Pure VLA
50.0R0
73.9R1
76.1R2
R0 Shared base skill
385 / 770 solvedR1 One reflection round
184 / 385 solvedR2 Two reflection rounds
17 / 201 solved
385 / 770 solvedR1 One reflection round
184 / 385 solvedR2 Two reflection rounds
17 / 201 solved
Inside the skill library
What reflection changes
06 / Integrated assistance
One request.
Both loops at work.
“Please serve me lunch.” One request brings personalization, protection, and recovery together.
Carrot runReal robot · 9×
Meat runReal robot · 7.5×
View the task sequence
Physical feedback guides both how to continue and whether the task should continue.
CarryOn
Personalized decisions.
Protected and recoverable execution.
Persistent context. Runtime protection. Recovery that improves with experience.