Figure 03 + Prismatic

1. Start with what Prismatic already has

Figure says Helix controls perception, movement and reasoning onboard and in real time, while its newer fleet infrastructure tracks robot health, location and operational status and pushes software/behaviors across the fleet. figure.ai

3. Figure's stack is now very sophisticated

A simplified version looks like this:

             Figure Fleet Platform
       Health • location • status • OTA
                    │
                    │
                  Helix
                    │
              System 2
       Language / scene reasoning
                    │
              System 1
        Visuomotor transformer
              ~200 Hz
                    │
              System 0
       Whole-body neural control
               ~1 kHz
                    │
               Figure 03
        Sensors • hands • body
                    │
              Physical world

That is a remarkably vertically integrated stack.

Figure controls almost everything from:

data → foundation model → reasoning → perception → motion → hardware → manufacturing.

figure.ai

4. Where Prismatic should sit

Above the fleet/robot intelligence layer:

                    PRISMATIC
         Persistent Cognitive Infrastructure

 Identity • Long-term memory • Experience
 Relationships • Knowledge • Evidence
 Decisions • Procedures • Governance
 Work history • Organizational memory
                     │
                    PPAI
                     │
             Figure Fleet Layer
                     │
                   Helix
                     │
          S2 → S1 → S0
                     │
                Figure 03
                     │
              Physical world

That gives us four different questions.

Figure 03

Can my body physically do this?

Helix

How do I perceive, reason and perform it?

Figure Fleet

Where are the robots, what are they doing, and what software should they run?

Prismatic

What has this machine experienced over its life, what does the organization know because of it, and what should remain true when the technology underneath changes?

That's the clean boundary.

5. Figure is especially strong in long-horizon task memory

This needs careful wording because Figure does have memory-like capability.

Helix 02 demonstrated a four-minute autonomous household task comprising 61 ordered loco-manipulation actions, maintaining task state and recovering from errors throughout the sequence. figure.ai

So we shouldn't claim:

“Figure has no memory.”

It clearly maintains working/task state.

But that's different from what you're building.

Helix task state

“I opened the dishwasher, moved the bowls, and still need to place the cups.”

Duration:

seconds/minutes.

Prismatic persistent memory

“Eight months ago this robot encountered a recurring dishwasher latch issue in this home, the owner approved a workaround, maintenance later replaced the latch, and the problem disappeared.”

Duration:

months/years.

That's the distinction.

6. Figure's fleet learning is formidable

Figure 03 can use 10 Gbps mmWave data offload to upload large amounts of robot data for continuous learning. figure.ai

Figure is therefore building something like:

Robot experience
     ↓
Raw physical data
     ↓
Fleet upload
     ↓
Training corpus
     ↓
Train improved Helix
     ↓
OTA deployment
     ↓
Entire fleet gets smarter

That's a very powerful loop.

But look at what happens to the individual experience.

The experience primarily becomes:

training data for improving the general model.

Prismatic's idea is different:

Robot experience
     ↓
Semantic event
     ↓
Evidence
     ↓
Individual memory
     ↓
Site/family memory
     ↓
Validated organizational knowledge

One improves the model.

The other preserves the experience.

That's a very important distinction.

7. Figure 02 → Figure 03 gives Prismatic a fantastic example

This is perhaps the most interesting thing in today's research.

Yesterday, September 30, 2026, Figure announced the decommissioning of its Figure 02 fleet as Figure 03 expands. figure.ai

That is almost a perfect illustration of Prismatic's thesis.

Figure 02 accumulated:

factory experience
deployments
failures
interactions
tasks
operational context

Then:

Figure 02
    ↓
retired
    ↓
Figure 03

Figure certainly uses what it learned to improve the next generation.

But imagine Prismatic:

Physical body:
Figure 02 → retired

New body:
Figure 03

Identity:
Worker-027

Memory:
preserved

Relationships:
preserved

Work history:
preserved

Procedures:
preserved

Decisions:
preserved

Evidence:
preserved

So instead of:

new robot generation = new machine

you could have:

new body = continuation of the same cognitive worker

That is one of the strongest illustrations we've found for Prismatic.

figure.ai

8. Imagine this at BMW

Figure 02 worked at BMW for roughly 1,250 hours, loaded more than 90,000 parts, and contributed to production of more than 30,000 vehicles before the generation was retired. figure.ai

Now Figure 03 has arrived at the same BMW Spartanburg plant and is performing a more complicated sequencing workflow using Helix 02. figure.ai

Today that transition looks roughly like:

Figure 02
   ↓
deployment data
   ↓
Figure engineering
   ↓
Figure 03

With Prismatic it could become:

Figure 02 body
      │
      │ experience
      ▼
 PRISMATIC IDENTITY
      │
      │ continuity
      ▼
Figure 03 body

Imagine Figure 03 arriving and already knowing:

“I worked this BMW facility previously.”

“Station 7 historically generated alignment problems.”

“These are the maintenance procedures approved by this plant.”

“These are the human operators I have worked with.”

“Here are the decisions made during the previous generation.”

That's not model training.

That's cognitive continuity.

9. Figure is also entering the home — which makes persistent relationships much more valuable

Figure now explicitly positions Figure 03 for household work such as:

laundry
cleaning
dishes

and says users can talk to the robot, ask it questions and delegate tasks. figure.ai

That makes Prismatic even more interesting.

Factory memory matters.

But household memory could be much richer.

Without Prismatic:

“Put the laundry away.”

Helix understands and executes.

With persistent cognition:

“Use the shelf arrangement we decided on last month.”

“Don't put these clothes in the dryer because Shazia prefers them air-dried.”

“You normally prepare the guest room before the children visit.”

That's where relationships and long-term contextual memory become product features rather than just industrial infrastructure.

11. Figure's multi-robot work is another interesting overlap

Figure has demonstrated two Helix-equipped robots coordinating physical tasks without central messaging or a common planner; each observes what the other is doing and adapts. figure.ai

Again, that's not the same as Prismatic.

Figure coordination

“What is the other robot doing right now?”

Prismatic organizational cognition

“What has this other robot learned over the last six months that I should know?”

One is:

real-time physical coordination.

The other is:

institutional knowledge sharing.

They're complementary.

A production-grade cognitive event adapter.

Something like:

ppaiclient-figure

It would consume safe higher-level events:

robot.started
task.started
task.completed
task.failed
skill.used
operator.interaction
battery.event
maintenance.event
hardware.changed
model.updated
deployment.changed
anomaly.detected

Then Prismatic creates:

experience.recorded
memory.updated
relationship.updated
evidence.linked
pattern.detected
procedure.updated
decision.requested
lesson.promoted

That's enough.

You don't need access to Figure's 1-kHz control loop.

14. Figure exposes another Prismatic feature I would formalize

I would add:

Cognitive Migration

Not simply backup/restore.

A formal operation:

EXPORT COGNITIVE IDENTITY
        │
        ▼
Identity
Relationships
Validated memories
Work history
Procedures
Evidence
Open commitments
Organizational membership
        │
        ▼
IMPORT INTO NEW BODY

Example:

Figure 02
    │
    │ cognitive migration
    ▼
Figure 03

Or someday:

Figure 03
    │
    ▼
Figure 04

Or even:

Figure
   │
   ▼
Unitree / Atlas / Apollo

provided the organization permits it.

Now hardware independence becomes something visible and demonstrable rather than just architectural language.

15. Figure vs the others we've studied