GR00T
GRASP
Generalist policy · specialist grasp

The Handoff

A generalist control model moves the arm; a specialist grasp planner decides how to pick the object up.

Watch the live demo
A generalist policy paired with a specialist grasp planner · GR00T1 × TRO-Grasp2 / Suction-Grasp3 · 2026
01The generalist

GR00T1 — a generalist model for many robots.

GR00T architecture: a camera image and a language instruction feed a vision-language model, then a diffusion transformer outputs motor actions
Vision + language in
Actions out, step by step
One model, many robots

But it learned standard two-finger grippers — not multi-finger hands or suction cups.

1 GR00T — NVIDIA, 2025 · arXiv:2503.14734
02The catch

A gripper can't hold a big object.

Grasp fails
WidowX + two-finger gripper · the apple is too wide to hold

A standard two-finger gripper can't wrap a large object like an apple — it never forms a stable hold, so the lift fails.

03A bigger hand

A bigger hand can hold it — GR00T can't drive it.

Grasp exists
TRO-Grasp2 · a stable 3-finger Barrett grasp on the apple
No retrain · fails
Same model, Barrett hand swapped in · GR00T can't control the new hand

The hardware can hold it. The generalist just never learned this hand — so control breaks the moment we change it.

2 TRO-Grasp — arXiv:2510.12724 · Barrett grasp visualization from the authors' demo
04The specialists

Models that focus only on the grasp.

Multi-finger Barrett hands grasping an apple, a can and other objects
Dexterous hand

TRO-Grasp2

A planner for a multi-finger Barrett hand — where to place each finger for a stable hold.

Predicted suction points and approach directions on a cluttered tabletop scene
Suction cup

Suction-Grasp3

A planner for a vacuum cup. Picks the best spot and angle to seal onto the object and lift it.

Each answers one question: given this object's shape, how should the hand make contact?

2 TRO-Grasp — arXiv:2510.12724·3 Suction-Grasp (SuctionNet-1B) — Cao et al., 2021 · arXiv:2103.12311
05The approach

Compose,
don't retrain.

Keep the generalist model unchanged. Bring in a specialist just for the grasp, then hand control back — a modular combination of generalist and specialist, with no extra training.

06The handoff

The model and the planner take turns.

 Take turns — pass control, then take it back
STEP 1 · MODEL

Reach

The control model moves the arm to the object.

STEP 2 · PLANNER

Grasp

Control passes to the planner, which sets the hand and secures the object.

STEP 3 · MODEL

Carry & place

Control returns to the model, which lifts, moves, and releases.

 The link — one shared message during the grasp
GR00T1
control model
a 3-D snapshot of the object
where and how to grip
Planner
grasp specialist

A small controller decides when to pass control — the model never has to learn the grasp itself.

1 GR00T — NVIDIA, 2025 · arXiv:2503.14734
07Generality

One interface. Any brain, any hand.

The link is just a standard message, so either side can be replaced on its own. TRO-Grasp2 and Suction-Grasp3 are only two instances — the same recipe accepts a new control model or a new end-effector, with no extra training.

Control model

the brain
GR00T1
Any vision-language policy
…any model that drives the arm

Grasp specialist

per end-effector
Parallel gripper
Dexterous hand — TRO-Grasp2
Suction cup — Suction-Grasp3
…any end-effector
1 GR00T — NVIDIA 2025, arXiv:2503.14734·2 TRO-Grasp — arXiv:2510.12724·3 Suction-Grasp (SuctionNet-1B) — Cao et al., 2021 · arXiv:2103.12311
08The result

Combine generalist & specialist

Success
WidowX · Barrett hand · apple on plate
Success
Panda · suction cup · red ball

Same recipe — a frozen generalist plus a grasp specialist — across two robots and two different end-effectors. No retraining either time.

09Related work

Where the handoff is new.

Approach Generalist,
language-driven
Specialist grasp
(dexterous + suction)
No
retraining
Many
robot arms
Any
end-effector
End-to-end VLAs
RT-2 · OpenVLA · π0 · GR00T
Dexterous-grasp VLA
DexGraspVLA
Grasp synthesizers
AnyGrasp · T(R,O) · SuctionNet
Modular VLM agents
MAESTRO · VoxPoser
The Handoff (ours)
GR00T × TRO-Grasp / Suction
yes partial no

Each prior approach covers part of the matrix; composing frozen modules over one standard interface is what covers all of it — generalist control and specialist grasps, across robots and end-effectors, with no retraining.

VLAs: RT-2 (2023) · OpenVLA (2024) · π0 (2024) · GR00T (2025)·DexGraspVLA (2025)·Grasp: AnyGrasp (2023) · Contact-GraspNet (2021) · T(R,O) (2025) · SuctionNet (2021)·Modular: VoxPoser (2023) · MAESTRO (2025)
Edit mode — click any text · Ctrl+S to save