# Can Better Representations Beat More Robot Data in VLA Training?

A new paper published today on arXiv delivers a pointed answer to one of the most practically important questions in humanoid AI development: when you cannot afford to collect more robot data, how far can smarter pre-training take you? The answer, according to researchers behind **VLAct**, is surprisingly far. On the RoboCasa-GR1 benchmark — an unseen humanoid embodiment the model was never explicitly trained on — VLAct using only **20% of downstream trajectories** outperforms the full-data GR00T-N1.6 baseline. The entire training run required a **16-GPU setup** using fully open-source data.

This is not a marginal improvement on a toy task. It is a direct head-to-head comparison on a humanoid platform, achieved at a fraction of the compute and data budget of an industrial system. The result reframes the dominant assumption that data volume is the primary lever for improving [Vision-Language-Action model](https://humanoidintel.ai/glossary/vision-language-action-model) performance.

The paper's core claim: **representation quality is a largely independent axis of VLA progress**, one that the field has underinvested in relative to data scaling.

---

## What VLAct Actually Does Differently

VLAct is described by its authors as a VLA-oriented VLM backbone trained on broad, heterogeneous, multi-embodiment robot data *before* task-specific fine-tuning. Three technical mechanisms distinguish it from straightforward continued pre-training:

1. **VLM-prior preservation** — prevents catastrophic forgetting of the broad visual-language understanding baked into the base VLM during robot-data adaptation.
2. **Multi-head continuous action co-supervision** — trains shared action semantics across embodiments without collapsing them into a single representation.
3. **Partially unified cross-embodiment action layout** — a structured approach to aligning action spaces that still allows task-specific action heads at fine-tuning time.

The architecture acknowledges a tension that most cross-embodiment VLA papers gloss over: you want shared representations for transfer, but you need embodiment-specific flexibility at deployment. The "partially unified" framing is VLAct's proposed middle ground, though the paper does not provide ablations granular enough to fully isolate which of the three mechanisms drives the largest gains — a gap worth flagging.

---

## The Benchmark Numbers

The paper reports results across several benchmarks, all grounded in the source text:

- **LIBERO-Plus**: VLAct achieves a success rate of **82.6%**, surpassing both ABot-M0 and LingBot-VLA.
- **RoboTwin 2.0**: VLAct achieves **92.5%** success rate, again beating the same industrial baselines.
- **RoboDojo**: VLAct ranks **sixth among all policies** by success rate and outperforms all explicitly designated world-action model (WAM) entries on both reported metrics.
- **RoboCasa-GR1 (unseen humanoid embodiment)**: VLAct with 20% of downstream data outperforms the full-data GR00T-N1.6 baseline.

The GR1 result is the headline because it directly tests [zero-shot generalization](https://humanoidintel.ai/glossary/zero-shot-generalization) to a humanoid embodiment not seen during VLAct's pre-training phase. This is the regime that matters most commercially: a policy that transfers efficiently to new hardware bodies reduces the per-platform data collection burden that currently makes humanoid deployment economics so difficult.

---

## Why This Matters for the Humanoid Stack

The robotics industry has implicitly organized around a data-centric scaling thesis. Large players collect proprietary teleoperation datasets; the hypothesis is that whoever accumulates the most embodied data wins. VLAct's results complicate that narrative in a specific, useful way.

**The data-scaling thesis is not wrong — it is incomplete.** Robot trajectories are inherently harder to scale than internet-scraped image-text pairs because physical collection is expensive, slow, and geographically sparse. A team that can extract more transferable knowledge from a fixed data budget has a structural efficiency advantage, especially at the current stage where even well-funded humanoid companies are working with tens of thousands of demonstrations rather than millions.

For companies building the AI stack for humanoids — [Physical Intelligence (π)](https://humanoidintel.ai/companies/physical-intelligence) and [Skild AI](https://humanoidintel.ai/companies/skild-ai) being the two best-capitalized pure-play examples — VLAct's framing of representation quality as an independent optimization axis is a direct invitation to audit their pre-training pipelines. The 16-GPU compute budget cited in the paper is also a deliberate provocation: this is a university-scale result, not a hyperscaler result.

The cross-embodiment transfer angle is particularly relevant for any company deploying multiple hardware generations or licensing their AI stack to third-party robot OEMs. A representation that generalizes efficiently across embodiments reduces the NRE cost of each new platform.

---

## Skeptical Read

A few caveats deserve emphasis before the community overcorrects.

First, the GR1 comparison is against GR00T-N1.6 specifically at a data-constrained operating point (20% vs. 100%). The paper does not report VLAct's performance when *also* given 100% of downstream trajectories, which would clarify whether this is a data-efficiency story or a representation-quality story that saturates quickly.

Second, benchmark selection matters. LIBERO-Plus, RoboTwin 2.0, RoboDojo, and RoboCasa are simulation environments. Real-world results are mentioned in the abstract as part of the evaluation scope, but the specific numbers cited in the paper are simulation benchmarks. [Sim-to-real transfer](https://humanoidintel.ai/glossary/sim-to-real-transfer) gaps remain notoriously hard to predict from simulation success rates alone.

Third, the paper's authors are affiliated with academic institutions, not the industrial labs whose systems they benchmark against. ABot-M0 and LingBot-VLA are named as industrial VLA systems, but no independent verification of those systems' training setups is available here. Benchmark parity assumptions deserve scrutiny.

None of these caveats invalidate the core result — they define its scope.

---

## Key Takeaways

- VLAct achieves **82.6%** on LIBERO-Plus and **92.5%** on RoboTwin 2.0, surpassing ABot-M0 and LingBot-VLA.
- On RoboCasa-GR1, VLAct with **20% of training data** outperforms the full-data GR00T-N1.6 baseline — the most commercially significant result.
- The method uses three mechanisms: VLM-prior preservation, multi-head continuous action co-supervision, and a partially unified cross-embodiment action layout.
- Training required only a **16-GPU setup** on fully open-source data — unusually accessible for research at this performance tier.
- Representation quality is framed as an independent axis of VLA progress, distinct from and complementary to data scaling.
- Results are primarily in simulation; real-world generalization claims should be evaluated with that context.
- The work directly challenges the assumption that proprietary large-scale robot datasets are a durable moat for AI stack companies.

---

## Frequently Asked Questions

**What is VLAct and how does it differ from standard VLA pre-training?**
VLAct is a VLA-oriented VLM backbone designed for continued pre-training on heterogeneous, multi-embodiment robot data before task-specific fine-tuning. It differs from standard approaches by explicitly preserving the original VLM prior, using multi-head continuous action co-supervision to build shared action semantics, and adopting a partially unified cross-embodiment action layout — all aimed at improving representation quality rather than simply fitting more action data.

**How does VLAct compare to Nvidia GR00T-N1.6 on humanoid tasks?**
On the RoboCasa-GR1 benchmark, which tests transfer to an unseen humanoid embodiment, VLAct trained on only 20% of downstream trajectories outperforms the GR00T-N1.6 baseline trained on the full dataset. This result is reported in simulation.

**What benchmarks did VLAct achieve top results on?**
VLAct achieved success rates of 82.6% on LIBERO-Plus and 92.5% on RoboTwin 2.0, surpassing industrial VLA systems ABot-M0 and LingBot-VLA on both. On RoboDojo it ranked sixth overall and outperformed all designated world-action model entries.

**Does VLAct require proprietary data or large compute?**
No. The paper explicitly states the training uses fully open-source data and a 16-GPU setup, positioning it as a compute-efficient alternative to data-heavy industrial VLA training pipelines.

**Why does cross-embodiment transfer matter for humanoid robotics?**
Humanoid hardware evolves rapidly and per-platform data collection is expensive. A policy that transfers efficiently to new embodiments — like VLAct's GR1 results suggest — reduces the data and time cost of deploying AI across successive hardware generations or multiple robot form factors.