ZETA: A Controlled Study of Zero-Shot Cross-Embodiment VLA Transfer for Tabletop Manipulation

1Galbot 2CFCS, School of CS, Peking University 3Peking University 4Renmin University of China 5The University of Hong Kong 6Xiamen University Malaysia 7Beijing Academy of Artificial Intelligence 8Beihang University
*Equal contribution. Corresponding authors.

CoRL 2026

Overview of ZETA's controlled pretraining, post-training, evaluation embodiments, and four research questions
Overview of our controlled study. We distinguish strict zero-shot transfer, where the target robot is absent from all training data, from pretrain-exposed zero-shot transfer. Under a controlled benchmark, we conduct fine-grained analyses across four embodiment shift categories, covering four research questions (RQ) on state-action representations, source-embodiment diversity, auxiliary co-training, and target-embodiment exposure.

Abstract

Zero-shot generalization to unseen embodiments is important for generalizable vision-language-action (VLA) models as robot hardware evolves and task-specific data collection remains costly. However, a systematic understanding of this problem remains limited, in part because the literature lacks a unified zero-shot transfer definition and controlled evaluation settings that isolate embodiment changes from differences in tasks, scenes, or protocols.

We distinguish strict zero-shot transfer, where the target embodiment is absent from all training data, from pretrain-exposed zero-shot transfer, where it appears only during pretraining. We introduce a controlled benchmark spanning 14 held-out target embodiments across simulation and real-world validation, then study state-action representations, pretraining embodiment diversity, auxiliary co-training objectives, and target-embodiment exposure.

Local end-effector state-action representations, the source-diversity comparison, and auxiliary co-training improve cross-embodiment transfer by around 15, 18, and 7 percentage points, respectively. Adding only 5% target-embodiment data during pretraining improves average target-embodiment progress by 13.4 percentage points, showing why strict and pretrain-exposed zero-shot transfer should be reported separately.

Controlled Benchmark

ZETA aligns tasks, object distributions, initial-pose regions, backgrounds, table layouts, lighting, and camera views between post-training and evaluation. The benchmark contains seven held-out simulation embodiments and seven held-out real-world embodiments, organized by the embodiment component that changes.

Four cross-embodiment evaluation categories in simulation and the real world
Appearance-only (APP), gripper-only (GRP), arm-only (ARM), and full-embodiment (FULL) shifts.

14

held-out target embodiments

6,300

simulation rollouts per model

140

real-world rollouts per model

4

controlled research questions

RQ1: State-Action Representation

We compare absolute and delta end-effector states with world-frame and end-effector-frame delta actions. Local EEF-centered representations preserve more transferable motion structure as the robot morphology changes.

StateActionSimulation averageReal-world average
Abs. EEFWorld-Delta60.356.0
Abs. EEFEEF-Delta64.661.6
EEF-DeltaWorld-Delta73.460.8
EEF-DeltaEEF-Delta75.789.9

Finding: EEF-Delta states and actions give the best average transfer in both simulation and real-world validation.

RQ2: Source-Embodiment Diversity

The main comparison fixes the total pretraining budget at 640K trajectories while varying the source pool from 1 to 8 to 512 embodiments. The 512-source setting outperforms the single-source setting, although the real-world 8-source result is non-monotonic as per-embodiment density falls before the source pool becomes sufficiently diverse.

RQ2 simulation and real-world results under a fixed total pretraining budget
Fixed-total-budget comparison and attention visualizations.

Complementary fixed-per-embodiment analysis

We also hold the trajectories per embodiment fixed while scaling from 32 to 128 to 512 sources. This complementary regime increases total data with source coverage; together, the two regimes provide complementary evidence rather than a complete causal isolation of diversity from data scale.

RQ2 fixed-per-embodiment simulation results
Simulation
RQ2 fixed-per-embodiment real-world results
Real world

RQ3: Auxiliary Co-Training

With the best RQ1 representation and the full 512-source pool, we test whether auxiliary objectives improve strict transfer beyond imitation-only training. Every evaluated co-training variant improves the average progress score.

Post-training objectiveAPPGRPARMFULLAverage
No co-training90.078.374.360.475.7
Language-action (EEF frame)88.686.779.370.581.3
Language-action (both frames)88.679.873.868.477.7
Subgoal90.387.477.871.581.8
Task-conditioned bounding box90.287.080.171.982.3

Finding: task-conditioned bounding-box co-training raises average progress from 75.7% to 82.3%.

RQ4: Target-Embodiment Exposure

RQ4 replaces a controlled fraction of the fixed 640K pretraining budget with target-embodiment data while keeping target downstream tasks excluded. Even limited exposure changes the transfer difficulty, motivating separate reporting for strict and pretrain-exposed zero-shot settings.

RQ4 target-embodiment exposure protocol
Target-exposure protocol
Target-exposure results for UR5eUMI
UR5eUMI
Target-exposure results for Google Robot
GoogleRobot

Finding: 5% target exposure improves UR5eUMI from 69.3% to 78.6% and GoogleRobot from 51.4% to 68.9%—13.4 percentage points on average.

RQ4 × RQ1: Representation and Target Exposure

Does target-embodiment exposure change which state-action representation transfers best? We compare all four RQ1 representations with 0% and 5% target-embodiment pretraining data. The total pretraining budget stays fixed, and target downstream tasks remain excluded.

Average task progress (%) on UR5eUMI and GoogleRobot.
StateAction0% target (strict)5% target (pretrain-exposed)
Abs. EEFWorld-Delta33.966.9
Abs. EEFEEF-Delta38.669.2
EEF-DeltaWorld-Delta58.467.5
EEF-DeltaEEF-Delta60.471.1

Finding: 5% target exposure improves all four representations and narrows the gap between the highest and lowest scores from 26.5 to 4.2 percentage points. EEF-Delta state and actions remain best, but representation choice matters most under strict transfer.

Qualitative Rollouts

RQ1: State-Action Representation

Real-world target embodiments

Franka Robotiq Source

Franka + CoRL sticker APP

Franka + UMI GRP

UR5e Robotiq ARM

UR5e + UMI FULL

Piper on quadruped FULL

Humanoid FULL

Humanoid + UMI FULL

Simulation target embodiments

Franka Source

Franka + green finger APP

Franka + CoRL sticker APP

Franka + UMI GRP

UR5e + Franka hand ARM

Google Robot + Franka hand ARM

UR5e + UMI FULL

Google Robot FULL

RQ2: Source-Embodiment Diversity

Simulation 1 source

Simulation 8 sources

Simulation 512 sources

Real world 1 source

Real world 8 sources

Real world 512 sources

RQ3: Auxiliary Co-Training

Language-action +5.6 pp

Subgoal +6.1 pp

Task-conditioned BBox +6.6 pp

RQ4: Target-Embodiment Exposure

Google Robot

Google Robot 0% target

Google Robot 5% target

Google Robot 30% target

Google Robot 100% target

UR5e + UMI

UR5e + UMI 0% target

UR5e + UMI 5% target

UR5e + UMI 30% target

UR5e + UMI 100% target

Practical Takeaways

  1. Report the zero-shot protocol explicitly. State whether the target embodiment is absent from every training stage or only from downstream post-training.
  2. Use embodiment-compatible representations. Local EEF-Delta states and actions are most reliable when the embodiment gap includes the arm or full robot.
  3. Cover diverse non-target embodiments. The 512-source setting gives the strongest arm-only and full-embodiment transfer, while the fixed-budget sweep remains non-monotonic in real-world validation.
  4. Use auxiliary supervision. Language-action, subgoal, and task-conditioned visual objectives all improve the imitation-only baseline.