StellarEdge AITechnical report · August 2026

Vision · Language · Action

StellaVLA

Learn the why, not just the motion.

In-Context Structured Demonstration for Generalizable Vision-Language-Action Models

StellaVLA adapts at test time from a single retrieved demonstration. Instead of replaying a trajectory, it transfers the task plan, subgoals, and grounded motion—without fine-tuning or autoregressive language decoding in the control loop.

Explore the method View the results

StellarEdge AI Technical Team

PaperCodeModelVLA-Arena
Overview of StellaVLA, showing human, XR and robot demonstrations transformed into a structured prompt for simulation and real-robot evaluation.
StellaVLA turns heterogeneous demonstrations into structured in-context guidance, then uses the same policy across simulation and real-world manipulation.
0.63#1 overall on VLA-Arena
98.8%Average success on LIBERO
85.1%Zero-shot on LIBERO-Plus
2.0×Faster with prefix caching

01 · The idea

A demonstration should explain the task.

Vision-Language-Action models can manipulate familiar objects, but their performance often collapses when the scene, viewpoint, or object changes. Collecting new data and fine-tuning for every new condition does not scale.

In-context imitation learning offers another path: retrieve one expert demonstration and condition the policy on it at test time. Yet raw observations and actions communicate only what an expert did. StellaVLA adds the missing why.

01

Conventional context

Raw trajectories preserve appearance—not intent.

Human, retargeted, and robot demonstrations all record a successful execution, but pixels and continuous actions alone do not explain the procedure that transfers between them.

Human, human-to-robot, and robot demonstration frame sequences.
Automatic offline structuringZero human annotation
A context demonstration organized into a task plan and subgoals containing an image, robot state, 2D trace, and 3D motion.
02

StellaVLA context

A compact explanation the policy can reuse.

Each keyframe is paired with the current subgoal, robot state, image-grounded 2D trace, and workspace-level 3D motion. The result is a task template instead of a trajectory replay.

Task planSubgoal image2D trace3D motion

“The policy learns to imitate the expert's reasoning, not only its motions.”

02 · Method

Structure offline. Learn in parallel. Act in real time.

A zero-annotation extraction pipeline and an asymmetric train/deploy design connect semantic reasoning to continuous robot control.

Automated subgoal context extraction from a task instruction and trajectory into subtask, robot state, 2D trace, and 3D motion.
01

Offline context extraction

Deduce the cause from the successful execution.

Qwen3-VL segments the trajectory into semantic subgoals. A deterministic verbalizer then projects robot motion into an image-grounded 2D trace and describes its 3D workspace displacement. No manual rationale labels are required.

Semantic SubtaskVisual 2D traceKinematic 3D motion
02

Parallel dual training

One representation, two complementary objectives.

The retrieved context, current observation, and task instruction enter a shared vision-language backbone. The language expert predicts the current subtask and grounded motion while the action expert regresses the continuous action chunk in parallel.

Language expert · training onlyAction expert · train + deploy
Shared vision-language model feeding a training-only language expert and an action expert in parallel.
03

Asymmetric inference

Remove language decoding. Cache the demonstration.

At deployment, StellaVLA keeps only the action path. The fixed demonstration prefix is encoded once and reused throughout the rollout, preserving structured guidance with approximately 2× faster context-conditioned inference.

Structured demoKV cache
Live observationVLM backbone
DeploymentAction expert

03 · Results

Generalization improves as the task gets harder.

StellaVLA preserves in-distribution competence while widening the gap under task, visual, and language shifts.

VLA-Arena overall0.63

First on the leaderboard as of August 1, 2026, compared with 0.44 for the strongest prior model reported in the technical report.

Task-level generalization

VLA-Arena overall score

Higher is better
0.84L0 · in distribution
0.62L1 · generalization
0.43L2 · hardest
Scenes from LIBERO, LIBERO-Plus, and VLA-Arena covering standard tasks, visual perturbations, safety, distractors, extrapolation, and long-horizon tasks.
Three simulation benchmarks probe in-distribution competence, zero-shot visual robustness, and task-level generalization.
LIBERO98.8%

Strong in distribution

Best average across Spatial, Object, Goal, and Long suites. The largest gains over the matched control appear on Goal and Long, where the current image alone cannot specify intent or order.

LIBERO-Plus85.1%

Robust without retraining

+10.1 percentage points over the matched demonstration-free control under camera, robot-state, sensor, layout, background, lighting, and language shifts.

Zero-shot robustness

Where structured context helps most

StellaVLA Matched control
Camera viewpoint70.5%
Robot initial state74.8%
Sensor noise92.8%
Language95.3%

04 · Evidence

The policy actually uses the demonstration.

Evaluation-time interventions reveal what carries the transferable signal—and when context can mislead the model.

Correct demonstration98.8%

Full four-suite LIBERO average.

No demonstration62.4%

The intended task becomes ambiguous.

Wrong-task demonstration44.9%

Mismatched context actively changes behavior.

98.8%

Text carries most of the structure

Text-only context nearly matches image + text on LIBERO, while image-only context falls to 92.9%. Structured language abstracts away source-specific appearance.

3

Sparse subgoals are already useful

Three keyframes reach 98.1%, within 0.7 points of the default ten. The subgoal sequence matters more than dense trajectory replay.

05 · Real world

One task language, across bodies.

Human-hand, XR-retargeted, and robot demonstrations share the same structured representation, while executable supervision remains in the target robot's native action space.

Source 01

Human hand

26 XR-recorded takes

Source 02

XR retargeting

26 frame-aligned robot trajectories

Target

AgileX Piper

Real-time 6-DOF control

125Teleoperated robot episodes
71,702Real-robot frames
85.0%In-distribution success
75.0%OOD-L1 success
Four AgileX Piper manipulation tasks and out-of-distribution variants: pen to cup, carrot to bowl, blocks into a drawer, and stacking bowls.
The real-robot benchmark spans precision placement, pick-and-place, articulated-object interaction, stacking, and controlled distribution shifts.
Real-robot success rates comparing StellaVLA, StarVLA-OFT and pi 0.5 across in-distribution and out-of-distribution tasks.
StellaVLA reaches 85.0% in distribution and 75.0% on OOD-L1. No model completes the unseen OOD-L2 task; StellaVLA records the strongest partial progress at 1.9 out of 4.

06 · Deployment

Reasoning learned once. Control executed fast.

Autoregressive rationale decoding is useful supervision, but it is incompatible with a high-frequency control loop. StellaVLA pays the reasoning cost during training, then deploys the action expert alone.

Context-conditioned inference2.0×faster with prefix caching
demo prefixcomputed oncelive observation

Technical report

StellaVLA

In-Context Structured Demonstration for Generalizable Vision-Language-Action Models

StellarEdge AI Technical TeamAugust 1, 2026Back to top ↑