InternVLA-A1.5

Unifying Understanding, Latent Foresight and
Action for Compositional Generalization

InternVLA-A1.5 Team

Paper Github Huggingface ModelScope BibTeX

Overview of InternVLA-A1.5, combining vision-language understanding, latent visual foresight, and action generation. — InternVLA-A1.5 uses learnable foresight tokens to query future-relevant latent information from shared multimodal context, while action queries decode continuous control through a lightweight expert.

98.9

LIBERO average

Best average success rate on the four LIBERO suites.

84.8

LIBERO-Plus total

Best total robustness under visual and layout perturbations.

93.2

RoboTwin average

Strong bimanual manipulation performance across clean and randomized splits.

861M

Training frames

Robot manipulation corpus spanning 1.2M episodes.

Introduction

A Quick Overview of InternVLA-A1.5

Start with the project video before the architecture walkthrough.

Model

Mixture-of-Transformers with training-only latent foresight

The model adopts a Mixture-of-Transformers design with a native VLM backbone and a smaller unified expert that receives multimodal context through shared full attention.

Tokenize the scene. Multi-view images, the language instruction, a control-mode token, and the 256-bin discretized robot state are embedded into one token sequence for the VLM.

Data

1.2M robot episodes, 861M frames, and 3M multimodal samples

A robot manipulation stream supplies action and future-frame supervision, while a multimodal stream preserves semantic and spatial-grounding ability.

Training data overview for InternVLA-A1.5. — Training data overview: six robot manipulation sources plus an InternVLA-M1 multimodal corpus for VQA and robotics-oriented grounding.

Source	Type	Episodes	Frames	Weight
InternData-A1	Sim.	587,946	395.9M	0.20
AgiBotWorld	Real	112,988	206.3M	0.25
UMI	Real	377,018	201.3M	0.10
DROID	Real	95,658	27.6M	0.15
Galaxea	Real	19,085	25.0M	0.20
RoboMind 1.0	Real	8,638	5.4M	0.10

Robot manipulation data mixture.

Source	Samples	Role
General QA	637K	Captioning, VQA, OCR, and knowledge grounding
Box QA	879K	Referring detection
Point QA	832K	Free-space and object-point localization
Trajectory QA	684K	End-effector waypoint prediction

Multimodal co-training data from InternVLA-M1.

Benchmarks

Broad simulation gains across static, robust, dynamic, and long-horizon benchmarks

The current paper reports simulation benchmark results across LIBERO, LIBERO-Plus, RoboTwin 2.0, SimplerEnv, DOMINO, and EBench.

Benchmark	Metric	InternVLA-A1.5
	Average SR	80.8
	Average SR	93.2
	SR / MS	27.7 / 39.8
	Average SR	98.9
	Average SR	84.8
	SR / Score	35.2 / 49.5

Click a benchmark row to view the full table.

Experiments

Instruction following and long-horizon chemistry tasks

The draft defines a four-task real-world suite but does not yet contain finalized success-rate numbers, so this page only describes the evaluated task setup.

Four real-world tasks for InternVLA-A1.5 evaluation. — The real-world suite includes Sort Tubes, Insert Tubes, Move Tubes, and MOF. Colored outlines mark task-relevant tubes and targets.
Sort Tubes
Instruction following with a specified tube and box target, including held-out arm-color bindings for compositional grounding.
Insert Tubes
The robot must bring a specified tube to a specified hole of a tube rack.
Move Tubes
The robot transports a specified tube to a specified position on the opposite side of the workspace.
MOF
A long-horizon chemistry procedure used to probe coherent multi-step execution.

Sort Tubes

Compositional Generalization: seen rollouts cover color and target factors; OOD rollouts swap them into held-out blue-right and orange-left instructions.

Seen: blue + leftSeen: orange + rightOOD: blue + rightOOD: orange + left

Seen:Blue tubeleft box

OOD:Blue tuberight box

OOD:Orange tubeleft box

Seen:Orange tuberight box

Insert Tubes

Compositional Generalization: the policy binds tube color to hole index and transfers that binding to held-out color-hole pairings.

Seen:BlueHole 1

OOD:BlueHole 2

Seen:BlueHole 3

Seen:BlueHole 4

Seen:OrangeHole 1

Seen:OrangeHole 2

Seen:OrangeHole 3

OOD:OrangeHole 4

Move Tubes

Compositional Generalization: the model reuses color and target-hole factors across a different manipulation primitive and OOD pairings.

OOD:BlueHole 1

Seen:BlueHole 2

OOD:BlueHole 3

Seen:BlueHole 4

Seen:OrangeHole 1

OOD:OrangeHole 2

Seen:OrangeHole 3

OOD:OrangeHole 4

Citation

BibTeX

@misc{internvla_a15,
  title        = {InternVLA-A1.5: Unifying Understanding, Latent Foresight, and Action for Compositional Generalization},
  author       = {InternVLA-A1.5 team},
  year         = {2026},
  howpublished = {\url{https://github.com/InternRobotics/InternVLA-A1.5}}
}