## Does CLAP Solve the VLM-to-VLA Transfer Problem?

A new paper from Yuri Ishitoya, Jeremy Siburian, Masashi Hamaya, Kuniaki Saito, Cristian C. Beltran-Hernandez, and Mai Nishimura proposes a deceptively simple fix to one of the hardest problems in robot policy learning: how do you convert a pretrained [vision-language-action model](https://humanoidintel.ai/glossary/vision-language-action-model) from a general-purpose VLM without destroying the semantic intelligence you're trying to inherit? Their answer — CLAP (Causal Language-Action Prediction) — achieves **90.8% on the LIBERO benchmark** with a 2B-parameter model after just single-epoch fine-tuning, a **+14.9 percentage point improvement over VLA-0**. The team will release weights at three scales — 0.8B, 2B, and 4B — all derived from a single VLM lineage, making this one of the more methodologically rigorous open-weight VLA releases the community has seen.

The core insight is architectural restraint. Rather than reshaping the backbone through large-scale post-training on robot data — the standard playbook that tends to bury whatever the VLM originally knew about language and semantics — CLAP keeps the backbone frozen in structure and prepends each numeric action sequence with a natural-language action description. This causal conditioning means the model generates a language plan first, then predicts precise action tokens anchored to that plan. No architectural modifications. The output-distribution mismatch problem — where predicting raw numeric action tokens pulls the model far from its pretrained language distribution — gets sidestepped by keeping the model in territory it already knows.

---

## The Output-Distribution Mismatch Problem, Explained

The fundamental tension in building [imitation learning](https://humanoidintel.ai/glossary/imitation-learning)-based VLAs from pretrained VLMs is that robot control requires generating bare numeric sequences — joint angles, end-effector deltas, gripper commands — that look nothing like natural language. Standard autoregressive language models are trained to predict the next token in a distribution built from human text. Shove action tokens into that distribution without adaptation and you degrade the model's language grounding. The more robot data you throw at it during post-training, the more you reshape the backbone — and the harder it becomes to understand whether downstream performance comes from the VLM's pretrained knowledge or from the robot data itself.

This interpretability gap matters more than it might seem. If teams building humanoid manipulation stacks can't isolate what a VLM actually contributes to control quality, scaling decisions become guesswork. You don't know whether to invest in bigger VLMs, better robot data, or architectural changes.

CLAP's approach — prepending natural-language action descriptions to numeric action sequences — keeps generation in the VLM's home territory. The causal structure means precise action-token prediction is conditioned on a language-action plan. The model "narrates" before it "acts," preserving the semantic structure the VLM was trained on.

---

## Benchmark Numbers and What They Mean

The reported numbers are specific and worth unpacking:

- **90.8% on LIBERO** (2B CLAP, single-epoch fine-tuning)
- **+14.9 percentage points** over VLA-0 at the same scale
- Improved robustness on **LIBERO-PRO** under three perturbation types: language, object, and spatial

LIBERO is a standard simulation benchmark for language-conditioned robot manipulation, so these numbers come with the usual sim-to-real caveats. Performance on LIBERO-PRO — a harder variant explicitly designed to stress-test robustness — is arguably the more meaningful signal here. The fact that CLAP shows gains under language perturbations (varied instruction phrasings) and spatial perturbations (shifted object positions) suggests the language-grounding mechanism is doing real work, not just overfitting to canonical task descriptions.

That said: simulation benchmarks are not factory floors. The paper's authors are releasing weights at 0.8B, 2B, and 4B, which will allow the broader community to probe real-world transfer. The multi-scale release from a single VLM lineage is particularly valuable — it enables controlled ablations on how capability scales with model size, something the field has lacked clean data on.

---

## Why This Architecture Matters for Humanoid Stacks

For teams building humanoid manipulation policies — whether for [dexterous manipulation](https://humanoidintel.ai/glossary/dexterous-manipulation) of small objects or [whole-body control](https://humanoidintel.ai/glossary/whole-body-control) tasks that require integrating language instructions with full-body motion — the CLAP framing addresses a real deployment constraint: you cannot afford to fine-tune massive backbone architectures every time you want to add a new task or adapt to a new environment.

CLAP's single-epoch fine-tuning result is the headline that hardware teams should notice. If policy adaptation can happen in one pass over robot data without backbone restructuring, the iteration cycle for humanoid task programming shortens dramatically. It also lowers the compute cost of per-deployment adaptation, which matters when you're scaling to fleets.

The open-weight release further democratizes access. Smaller humanoid startups that can't run proprietary VLA training pipelines at the scale of [Physical Intelligence (π)](https://humanoidintel.ai/companies/physical-intelligence) or [Skild AI](https://humanoidintel.ai/companies/skild-ai) now have a clean multi-scale baseline to build from — one that was designed from the ground up for interpretability and controlled comparison.

---

## Skeptical Take

A few caveats worth noting. First, the LIBERO benchmark, while useful, operates in simulation. [Sim-to-real transfer](https://humanoidintel.ai/glossary/sim-to-real-transfer) for language-conditioned manipulation remains an open problem, and no real-robot results are reported in the abstract. Second, the "+14.9 pt over VLA-0" comparison is useful, but VLA-0 is one specific baseline — the paper's significance will depend on how CLAP compares against a broader set of recent VLA approaches once the full paper is available for review. Third, the claim that single-epoch fine-tuning is sufficient is striking and will need replication across more diverse task distributions before the community treats it as a reliable property rather than a benchmark-specific artifact.

The multi-scale open-weight release is the real hedge here: by releasing 0.8B, 2B, and 4B variants, the authors are inviting exactly the kind of independent scrutiny these claims require.

---

## Key Takeaways

- **CLAP (Causal Language-Action Prediction)** converts pretrained VLMs into VLAs with no backbone architectural changes, addressing output-distribution mismatch by prepending language-action descriptions to numeric action sequences.
- **90.8% on LIBERO** with a 2B model after single-epoch fine-tuning — **+14.9 percentage points** over VLA-0.
- Robustness gains on **LIBERO-PRO** under language, object, and spatial perturbations suggest genuine semantic grounding, not benchmark overfitting.
- Open weights to be released at **0.8B, 2B, and 4B** from a single VLM lineage — enabling controlled, multi-scale analysis of VLM-to-VLA capability transfer.
- All results are currently in simulation; real-robot validation is the critical next test.
- Single-epoch fine-tuning, if it holds across task distributions, significantly reduces the compute cost of adapting policies for humanoid deployment.

---

## Frequently Asked Questions

**What is CLAP in robotics AI?**
CLAP (Causal Language-Action Prediction) is a method for converting pretrained vision-language models (VLMs) directly into vision-language-action models (VLAs) without modifying the backbone architecture. It prepends natural-language action descriptions to numeric action token sequences, causally conditioning action prediction on a language plan and preserving the VLM's pretrained semantic capabilities.

**How does CLAP compare to other VLA training approaches?**
Most VLA training involves large-scale post-training on robot data that reshapes the model backbone extensively. CLAP explicitly avoids this, using single-epoch fine-tuning and no architectural modifications. On the LIBERO benchmark, the 2B CLAP model scores 90.8%, which the paper reports as +14.9 percentage points over VLA-0 at the same scale.

**What is the LIBERO benchmark?**
LIBERO is a simulation benchmark for evaluating language-conditioned robot manipulation policies. LIBERO-PRO is a harder variant designed to test robustness under language, object, and spatial perturbations.

**Will CLAP weights be publicly available?**
According to the paper, the authors plan to release CLAP as an open-weight model family at 0.8B, 2B, and 4B parameter scales, all derived from a single VLM lineage.

**What does CLAP mean for humanoid robot manipulation?**
If single-epoch fine-tuning results generalize beyond simulation benchmarks, CLAP could significantly reduce the compute and data cost of adapting manipulation policies for new tasks on humanoid platforms — a meaningful practical advantage for teams operating at deployment scale.