AAAI 2027  /  Robot Learning Anonymous Submission   01
Training-free composition

GraspBridge

Zero-Shot Cross-Embodiment Bridging Between
Vision–Language–Action Policies and Specialist Grasp Planners

A frozen VLA keeps the task. Only the failed grasp is delegated to a frozen specialist planner — then control comes straight back.

Three arms — WidowX, Google Robot and Franka Panda — each shown with suction, Shadow, Barrett and LEAP end-effectors, every pairing marked as working
3
VLA families
3
Robot arms
4
End-effectors
0
Gradient steps
Policy
Plan
Execute
Transition
Policy
01 · MotivationThe generalist   02
The generalist

GR00T1 — a generalist model for many robots.

A camera image and a language instruction feed a vision-language model, then a diffusion transformer outputs motor actions
πθ( a  |  o, ℓ )

A camera observation o and a language instruction ℓ map to a short-horizon action chunk a. The parameters θ stay frozen throughout this work.

InVision + language
OutActions, step by step
ScopeOne model, many robots

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

1 · GR00T — NVIDIA 2025, arXiv:2503.14734Frozen πθ throughout
01 · MotivationWhere episodes break   03
The research gap

Reaching is reasoning.
Grasping is contact.

Reaching, transporting and placing are task-level problems the policy handles well. Grasping is a contact-rich control problem with entirely different physical constraints — and it is where generalist policies fail.

Grasp fails
WidowX · native two-finger jaw
the apple is too wide to hold
Reached
≈100%
→
Aligned
≈100%
→
Lifted
0%
→
Carried
—
→
Placed
—

Nested phase indicators on failed rollouts with a swapped end-effector. The policy reliably arrives at the object; the episode dies at contact.

MotivationPhase decomposition of a pick-and-place rollout
01 · MotivationThe naive swap baseline   04
Cross-embodiment deployment

Change the hand and the grasp collapses.

A frozen policy driving a WidowX arm succeeds with its native jaw but fails with suction, Shadow, Barrett and LEAP hands
The scalar gripper command is mapped straight to hand closure. Policy, tasks and seeds unchanged.
Naive swap · fails
WidowX · Barrett hand grafted in · the policy cannot drive it
≈100%
Still reach the object — the episode dies at contact
Baseline40 grafted arm × hand × task settings
01 · MotivationGrasp synthesis   05
The grasp specialists

High-quality grasps for exactly these hands already exist.

Grasp planners take an object point cloud and a hand model and return a ranked wrist pose, joint configuration and score. They are task-agnostic — geometry only, no language — and they run as standalone systems, disconnected from any policy.

Barrett and dexterous hands grasping an apple, a can and other objects
TRO-Grasp2 — learned graph diffusion

Ranked candidates for Barrett, Shadow and LEAP from a partial cloud.

Interactive · drag A Barrett three-finger hand closed around an apple
One planned grasp2 — Barrett on the apple

The planner's actual output: a wrist pose plus a joint configuration.

Predicted suction points and approach directions on a cluttered tabletop
SuctionNet4 — surface-normal contacts

Seal points and approach angles for the suction gripper.

We evaluate three: TRO-Grasp2 (learned, dexterous, from a partial cloud), BODex3 (pre-synthesised tabletop grasps for the same hands, aligned online to the observed pose) and a suction planner4.

2 · T(R,O) Grasp — Fei et al., arXiv:2510.12724  ·  3 · BODex — Chen, Ke & Wang, ICRA 2025  ·  4 · SuctionNet-1Billion — Cao et al., RA-L 2021𝒢(𝒫, h) → (Twrist, qh, s)
02 · MethodGraspBridge   06
The proposal

Compose at inference time. Retrain nothing.

GraspBridge pipeline: the instruction drives a frozen policy model; on a grasp signal a VLM detects the object, depth is back-projected into a point cloud, and a grasp planner returns an effector action
MODULE A

Frozen VLA policy

Owns the language, the reach, the transport and the placement.

MODULE B

Frozen grasp specialist

Invoked only on a grasp signal. Sees geometry, never the instruction.

THE INTERFACE

One service call

No shared weights. No co-training. Neither module is conditioned on the other.

Method overviewNeither module is retrained or fine-tuned
02 · MethodThe runtime handoff   07
Failure-triggered · phase-scoped

Four states, and control always comes back.

STATE 1 — POLICY

Forward, unmodified

VLA actions pass through untouched, while a detector watches for a failed grasp.

STATE 2 — PLAN

Build cloud, query planner

Reconstruct the object from the camera, and ask the specialist for a grasp.

STATE 3 — EXECUTE

Realise and verify

Drive the hand onto the grasp, close, and verify the lift — or retry.

STATE 4 — TRANSITION

Hand back in-distribution

Move to a canonical pose, so the policy resumes on familiar ground.

30–50
Ticks under specialist control
200–400
Ticks in the episode
1
Manipulation phase touched
Policy
Plan
Execute
Transition
Policy
02 · MethodThe loop in motion   08
The four states of the previous slide, running

Pass control, secure the object, hand it back.

Task success
State 1Reach
→
State 2Plan
→
State 3Execute
→
State 4Return
WidowX · Barrett · apple → plate
Task success
State 1Reach
→
State 2Plan
→
State 3Execute
→
State 4Return
Panda · suction cup · apple → basket

The same four states run in both clips: control passes to the specialist at State 2, the object is secured and verified in State 3, and State 4 hands the arm back in-distribution. Contact mechanics differ completely between these lanes — the wrapper and the interface do not.

Method · qualitativeStep track follows the clip in real time
03 · ExperimentsSetup   09
One wrapper, held fixed

Three arms. Four hands. Two simulators.

Frozen VLA policies
GR00T-N1.7-3BOcto-smallOpenVLA-7B
Arms
WidowX 250 S · 6 DoFGoogle Robot · 7 DoFFranka Panda · 7 DoF
End-effectors — none seen in any policy's training data
SuctionBarrett BH-282 · 3 fingerShadow · 24 DoFLEAP · 16 DoF
Grasp planners
TRO-GraspBODexSuction planner
Benchmarks · metrics
SimplerEnv · SAPIENLIBERO · MuJoCo6 tasks5 × 50 episodes

Baselines: the naive end-effector swap (identical policy, tasks and seeds) and each policy's native parallel jaw as a reference. We report task success and grasp success — the latter localises whether a failure belongs to the specialist or to the downstream policy.

Three arms — WidowX, Google Robot and Franka Panda — each shown with suction, Shadow, Barrett and LEAP end-effectors
ExperimentsWrapper, planners and hyper-parameters identical in every cell
03 · ExperimentsMain result   10
Task success (%) · naive swap vs +GraspBridge

GR00T-N1.7-3B results.

Native Suction Barrett Shadow LEAP
ArmTask jaw naive+Bridge naive+Bridge naive+Bridge naive+Bridge
WidowXapple → plate15.2 0.096.4 22.890.0 0.091.2 0.093.2
WidowXcube stack28.0 0.033.6 0.035.2 0.032.0 0.032.4
Googleapple → can34.8 0.0100.0 49.279.6 0.099.2 7.698.0
Pandared ball → basket8.8 0.0100.0 0.066.8 0.064.4 0.090.0
Pandamilk → basket90.4 0.099.6 62.888.4 0.089.6 0.090.8

Averaged over five runs of 50 episodes. cube stack is the exception, and not a grasp failure: it demands a precise release above a second object, so it is placement-limited.

+89.8
Median gain, this table (pts)
0
Lines of policy code changed
Table 1 · frozen GR00T-N1.7-3BFull grid also covers Octo and OpenVLA
04 · AnalysisPolicy-agnostic   11
Task success (%) · three frozen VLA policies

Results across policies.

PolicyShared laneNative jaw Barrett
+Bridge		Shadow
+Bridge		LEAP
+Bridge		Suction
+Bridge
GR00TWidowX · apple → plate15.2 90.091.293.296.4
OctoWidowX · apple → plate0.0 96.894.495.297.6
OpenVLAWidowX · apple → plate0.0 90.094.896.498.8
GR00TGoogle · apple → can34.8 79.699.298.0100.0
OctoGoogle · apple → can22.4 76.496.895.298.4
OpenVLAGoogle · apple → can18.8 74.895.294.097.6

The three policies differ in architecture, action parameterisation and training mixture. No policy-specific code is introduced.

OpenVLA on WidowX, rubber duck to plate: frozen policy versus policy plus GraspBridge, for the native jaw and the Barrett hand
OpenVLA-7B · WidowX · rubber duck → plate
AnalysisSame wrapper · same planners · same hyper-parameters
ConclusionGraspBridge   12
Takeaways

VLA failures are not uniform
across an episode.

Task reasoning, navigation and transport stay reliable under an end-effector swap. Only contact does not — so delegate only that phase.

1

A training-free, phase-scoped handoff between two frozen modules

Neither retrained, fine-tuned, nor conditioned on the other. One unchanged interface across three VLA families, three arms and four end-effector morphologies.

2

A grasp pose is not a grasp

The execution layer is the difference between a valid plan and a realised grasp — quantified mechanism by mechanism, and embodiment-dependent.

3

Deployable from RGB-D alone, and dispatchable before it runs

Both the trigger and the planner input come from observations a real robot has; a pure-Python router decides which arm–hand–task lane is admissible before any process starts.

WidowX · apple → plate0 → 96
Google · apple → can0 → 100
Panda · red ball0 → 100
Task success (%) · naive swap → +GraspBridge
Best end-effector per lane · frozen GR00T
Policy
Plan
Execute
Transition
Policy
GraspBridgeOne interface   13
The whole system

One wrapper. Three arms, four end-effectors.

Top left: a frozen policy succeeds on its native jaw and fails on suction, Shadow, Barrett and LEAP. Bottom left: GraspBridge grounds the object, back-projects a point cloud and returns an effector action while the policy keeps the arm. Right: all three arms working with all four end-effectors.
Naive swap · the GraspBridge loop · universal adaptation
DemoLive demo   14
Live in the browser · nothing to install

An interactive results explorer.

Pick a policy, an arm, a hand and a task — the selected run is then shown three ways.

VIEW 1 · RESULTS

The whole grid

The headline table as charts, and as the full grid behind them.

VIEW 2 · COMPARISON

Three lanes at once

Native gripper, bare swap and grasp specialist on one task, side by side.

VIEW 3 · VIDEO

One episode, phase by phase

A rollout under a timeline coloured by who holds control at each tick.

That timeline is the rail of this talk — base policy driving, planned grasp, handoff / resume — so the claim of the paper can be checked run by run, rather than taken on the tables alone.

QR code linking to the GraspBridge interactive demo
crest-demo-handoff.pages.dev/demo/

Same runs, same checkpoints and same seeds as the tables in this talk.

Policy
Plan
Execute
Transition
Policy
Edit mode — click any text · Ctrl+S to save