# Can Behavior-Aligned Representations Solve Cross-Embodiment Transfer in VLA Models?

A Stanford research team has identified a concrete, quantified answer to one of humanoid robotics' most persistent data problems: how to make manipulation policies trained on one robot body work on another. Their finding — that [end-effector](https://humanoidintel.ai/glossary/end-effector) traces as an intermediate representation improve sim-to-real cross-embodiment transfer, boosting task completion progress by **28%** on real robot policies pre-trained on simulation data — gives the field a practical handle on a problem that has resisted clean solutions for years.

The paper, "Cross-Embodiment Transfer via Behavior-Aligned Representations" (arXiv:2607.27549), comes from Ajay Sridhar, Jensen Gao, Jonathan Yang, Jean Mercat, Suneel Belkhale, and Dorsa Sadigh. It introduces a simulation-based benchmark designed specifically to evaluate cross-embodiment transfer across diverse robot datasets to new, previously unseen embodiments — filling a methodological gap that has made apples-to-apples comparison of transfer approaches nearly impossible until now.

The 28% improvement in task completion progress on real hardware is the headline, but the paper's structural contributions — a rigorous benchmark, a comparative analysis of representation types, and evidence that these representations unlock value from action-free data — are arguably more durable for the industry.

---

## What Are Behavior-Aligned Representations, Exactly?

The core hypothesis is elegant: if a representation captures *what the robot is doing* in a way that is largely independent of *which robot is doing it*, then it should serve as a bridge across embodiments in a [vision-language-action model](https://humanoidintel.ai/glossary/vision-language-action-model).

The authors test three candidate representations:

- **Object bounding boxes** — spatial location of task-relevant objects in the visual frame
- **Language motions** — natural language descriptions of robot motion
- **End-effector traces** — the spatial trajectory of the robot's end effector over time

The intuition is that all three possess a degree of *embodiment invariance*: a cup is a cup, a "move left" description is a "move left" description, and an end-effector arc through space looks similar whether the arm is a Franka, a UR5, or a custom humanoid arm — regardless of the underlying joint configuration that produced it.

This stands in contrast to raw action representations, which are deeply embodiment-specific: joint angles, torques, and motor commands are meaningless across robots with different kinematics, actuator counts, or [degrees of freedom](https://humanoidintel.ai/glossary/degrees-of-freedom).

The authors frame these representations as needing to satisfy two conditions simultaneously: invariance across embodiments, and predictiveness of robot actions. That dual requirement is what separates a useful bridging representation from a merely generic one.

---

## What the Benchmark Actually Tests

One underappreciated contribution here is the benchmark itself. Cross-embodiment transfer research has historically suffered from heterogeneous evaluation setups that make results incomparable across papers. This team built a simulation-based benchmark explicitly designed to assess transfer from diverse, large-scale cross-embodiment training data to *new* embodiments not seen during training.

This is the harder, more realistic evaluation regime. It is not asking whether a policy trained on embodiment A transfers to embodiment A with a different payload. It is asking whether representations learned across a broad corpus of robot data generalize to a genuinely novel morphology — the scenario that matters for humanoid OEMs deploying new hardware variants or for foundation model providers like [Skild AI](https://humanoidintel.ai/companies/skild-ai) or [Physical Intelligence (π)](https://humanoidintel.ai/companies/physical-intelligence) who must serve diverse customer hardware.

---

## Three Findings That Matter for Humanoid Deployment

### 1. End-effector traces outperform other representation types

Among the three representation types tested, end-effector traces were identified as "particularly beneficial" for transfer. This is not entirely surprising — a Cartesian trajectory through task space is arguably the most morphology-agnostic description of manipulation behavior available without stepping all the way up to task-level semantics. But having it confirmed empirically, with a benchmark to back it up, is meaningful.

For humanoid arm developers, this suggests that trajectory-level supervision signals — rather than raw joint-space action labels — may be the right currency for building cross-embodiment datasets. It also has implications for teleoperation system design: capturing end-effector pose streams may be more transferable than capturing joint encoder data, even when the latter is easier to log.

### 2. Representations compound with dataset scale

The paper finds that behavior-aligned representations "are generally more useful with larger prior datasets." This is a critical scaling result. It implies that the approach is not a niche trick that works only in constrained lab settings — it becomes *more* valuable as the training corpus grows. For the humanoid industry, which is actively racing to aggregate large-scale manipulation datasets (Open-X Embodiment and its successors being the most prominent examples), this is a strong argument for building behavior-aligned annotation pipelines into data collection infrastructure from the start, rather than retrofitting them later.

### 3. Unlocking action-free data

Perhaps the most commercially significant finding: behavior-aligned representations enable [imitation learning](https://humanoidintel.ai/glossary/imitation-learning) to benefit from *action-free* data — video demonstrations where no robot action labels are recorded. This dramatically expands the usable data universe. Human manipulation videos, instructional footage, and existing video datasets that lack robot teleoperation logs all become potentially useful training signal.

For teams trying to scale data without proportionally scaling teleoperation costs — a very real constraint for every humanoid startup right now — this is a meaningful result. It suggests a path toward training on internet-scale video with embodiment-specific behavior-aligned representations as the bridging layer.

---

## Sim-to-Real: The 28% Number in Context

The 28% improvement in task completion progress on real robots — pre-trained on simulation data — is the result that will get the most attention, and it deserves some scrutiny.

The source abstract describes this as improvement in "task completion progress," not task success rate. These are related but distinct metrics. Task completion progress typically measures how far through a task sequence a robot gets before failing, rather than whether it finishes. This is a softer metric, and appropriately so for [sim-to-real transfer](https://humanoidintel.ai/glossary/sim-to-real-transfer) evaluation where zero-shot generalization from sim to real hardware is genuinely hard. A 28% improvement in progress is a real and useful gain, but readers should not conflate it with a 28% lift in binary task success.

That said, any measurable, consistent improvement in sim-to-real cross-embodiment transfer is significant. The sim-to-real gap for manipulation — as opposed to locomotion, where sim-to-real has matured considerably — remains one of the harder open problems in the field. Humanoid developers doing pre-training in simulation before deploying on physical platforms (a near-universal strategy given hardware cost) have a direct interest in any technique that narrows this gap.

---

## Industry Implications

This research lands at a moment when the humanoid industry's data strategy is crystallizing. Multiple companies are building proprietary teleoperation fleets, negotiating data-sharing arrangements, and evaluating foundation model providers. The question of how to make data collected on one robot body useful for another — whether across hardware generations, across OEM partnerships, or across sim and real domains — is not academic. It is a core strategic variable.

The behavior-aligned representation framework suggests a concrete architectural choice: build VLA models with intermediate representations that are Cartesian-trajectory-centric rather than joint-space-centric. Teams at companies building general-purpose manipulation stacks — from [Agility Robotics](https://humanoidintel.ai/companies/agility-robotics) to [Figure AI](https://humanoidintel.ai/companies/figure-ai) to [1X Technologies](https://humanoidintel.ai/companies/1x-technologies) — face this design decision in their data pipelines and model architectures today.

The benchmark contribution also matters beyond this paper's specific results. A standardized, simulation-based cross-embodiment evaluation framework gives the field a shared measuring stick. That is infrastructure the research community has needed.

---

## Key Takeaways

- **28% improvement** in task completion progress on real robots using behavior-aligned representations for sim-to-real cross-embodiment transfer, as reported by Sridhar et al. (Stanford, arXiv:2607.27549)
- **End-effector traces** outperform object bounding boxes and language motions as bridging representations in VLA models across embodiments
- **Scaling effect confirmed**: behavior-aligned representations provide greater benefit as the training dataset grows — directly relevant to large-scale data aggregation efforts
- **Action-free data** becomes usable training signal when behavior-aligned representations are incorporated — a potential multiplier on effective dataset size
- A **simulation-based cross-embodiment benchmark** is introduced, providing a standardized evaluation framework the field has lacked
- The 28% metric measures **task completion progress**, not binary success rate — an important distinction for interpreting the magnitude of the result

---

## Frequently Asked Questions

**What is cross-embodiment transfer in robotics?**
Cross-embodiment transfer is the ability of a robot policy — a learned behavior model — to operate effectively on a robot with a different physical body (morphology, actuator configuration, or kinematic chain) than the one it was primarily trained on. It is a central challenge for building general-purpose manipulation AI that can serve diverse hardware platforms without retraining from scratch for each one.

**What are behavior-aligned representations in VLA models?**
Behavior-aligned representations are intermediate prediction targets in vision-language-action models that describe *what is happening* in a manipulation task in a form that is largely independent of the specific robot performing the task. Examples include end-effector traces (the spatial path of the robot's hand), object bounding boxes, and language descriptions of motion. They sit between raw visual input and robot-specific action outputs, providing a morphology-agnostic bottleneck.

**Why are end-effector traces more transferable than joint angles?**
End-effector traces describe motion in Cartesian task space — the actual path the robot's hand takes through the world — rather than in joint space, which encodes the specific configuration of a particular robot's linkages and actuators. Because the Cartesian trajectory of a manipulation action is similar across robots performing the same task (pick up a cup, place an object), it generalizes across embodiments far better than joint-level commands, which are entirely specific to each robot's mechanical design.

**How does this approach relate to using video data for robot training?**
By enabling policies to learn from action-free data (video without robot action labels), behavior-aligned representations open a path to training on large video corpora of human or robot manipulation that lack teleoperation logs. This is significant because action-labeled robot data is expensive to collect, while video data is abundant. The representation acts as a bridge between unlabeled visual demonstrations and robot-executable actions.

**What does this mean for humanoid companies building foundation models?**
Teams building general-purpose manipulation foundation models — intended to serve multiple hardware platforms — should consider incorporating end-effector trace prediction as an intermediate objective in their VLA architectures. The evidence suggests this architectural choice improves cross-embodiment generalization, scales favorably with dataset size, and expands the usable training data universe to include action-free sources. It is a concrete, empirically-grounded design signal in an area where architectural choices have historically been made with limited comparative evidence.