Banseok Lee
CEO
Inference Expert
UCLA
Super Tiny
Robot Brain
Intelligence that fits inside the robot.
18 ROLLOUTS / 6× PLAYBACK
Robot brains are getting better
The next challenge is deployment.
Two ways to deploy.
Two problems.
01 / CLOUDRemote brain
NETWORK
RobotConnection risk
02 / ON BOARDJetson Thor
LOCAL COMPUTE
RobotCost + power
Both become harder at fleet scale.
Customer-required task success rate
99.997%
A dropped connection
can’t mean a failed task.
NETWORK OFFLINE LOCAL FALLBACK
The onboard trade-off
$5,000
Jetson Thor T5000 compute module*
Jetson Thor T5000 compute module*3h work / 1h rest
Operating scenario
Downtime eats ROI.
*NVIDIA module MSRP: $4,999 at 1,000+ units. Duty cycle depends on the robot and workload.
OUR GOAL
$1,000
Jetson Orin NX 16 GB compute module*
Built to keep
working.
Intelligence, fully on board.
LOCAL INFERENCE
ON BOARDSmall compute. Continuous inference.
*Module MSRP: $999 at 1,000+ units. Continuous operation is the design goal.
MODEL COMPRESSION · LittleBit (NeurIPS ’25 · ICML ’26)
100×
smaller language models
1 GB~10 MB
DEVELOPED ATSamsung Research
GALAXY DEPLOYMENT PIPELINE200M devices per year
LittleBit + Speculative Decoding
2× faster
LittleBit + World Action Model
160×
compressed.
Still working.
UnifoLM-WLA-1.0
Jetson Orin NX · All inference on board
We have an
early signal.
4×PLAYBACK SPEED
60/90
Successful real-world trials
Zero-shot in new environments3 locations3 fruits2 basketsEarly checkpoint: <10% of planned training
Rollouts at 4×. 60/90 = 66.7%. One-minute trials with retries. The task was seen in training; evaluation setups were new.
OUR GOAL
The whole brain.
One small chip.
System 2VLM
System 1VLA · WAM
System 0Control
ALL ON BOARDJetson
Orin NX
Orin NX
01
NEXTSuper-Efficient
Binary Inference
Chip
Binary Inference
Chip
OUR AMBITION
Impossible
to hack.
The world's most secure robot.
01
No network connection.
Intelligence stays on board.
02
NEXT
Homomorphic encryption.
Compute on encrypted data.
NX✦
ON-DEVICE INTELLIGENCEJetson Orin NX
TEAM
Four researchers who worked
together at Samsung Research
Lohith Ravuru
CTO
Training Expert
IIT Bombay - JEE Rank 4
Youngmin Kim, PhD
CRO
Humanoid Expert
KAIST
Jungwoo Lee
CPO
AI & Robot Systems Expert
Yonsei University
The Ask
$20M
GPU compute
Model training & delivery
Research
$10M
Team
US HQ · 5
Korea R&D · 15
$5M
Hardware
Robots, sensors
& test rigs
$2M
SG&A
Office, travel
& operations
$3M
Milestones after funding
3 months
GR00T 2.0 + Alpamayo
On-device inference targets
6 months
First shipment to a
high-volume manufacturerOur model + runtime. Their data + hardware.
18 months
Fully Autonomous Robot
On-device reasoning + encryption
Three partners. One turnkey stack.
Goal: ship to 3 partners (1 Korea, 2 US) and build a turnkey solution.
World Action Model Humanoids, bi-arm
In discussion · LG
Target accounts · Figure, Tesla, 1X
VLA AMRs
In discussion · NAVER Labs
Target accounts · MiR, OTTO
On-Device Models
In discussion · Amazon
Target channels · Advantech, Seeed Studio
ModelRuntimeIntegration