## Does Synthetic TAMP Data Actually Help Fine-Tune VLA Models?

**The answer, according to new research from Samrat Sahoo, Yixuan Huang, and Tom Silver, is: only if you align it first.** Raw task and motion planning trajectories achieve just an 8.3% average success rate when used to fine-tune a pretrained [vision-language-action model](https://humanoidintel.ai/glossary/vision-language-action-model) — even though those same trajectories successfully solve the target tasks. The paper, released today on arXiv (2609.12316), introduces DATAFARM, a framework that aligns planner-generated trajectories with the VLA's pretraining distribution across robot joint configurations, motion style, and temporal execution profiles. The result: 56.7% average success rate, approaching the human teleoperation ceiling of 61.7% on the same task set. That is a 6.8× improvement over the raw TAMP baseline — and it arrives without collecting a single additional human demonstration.

This finding matters enormously for the humanoid industry. The field's central bottleneck is not actuator cost or walking stability anymore — it is data. Every major platform deploying [imitation learning](https://humanoidintel.ai/glossary/imitation-learning)-based manipulation needs demonstration data at scale, and TAMP has long been the theoretically attractive but practically underperforming answer to that problem. DATAFARM reframes why TAMP has underperformed and offers a concrete fix.

---

## The Distribution Mismatch Problem TAMP Has Always Had

TAMP planners optimize for task success, not for behavioral fidelity to any particular dataset. A planner solving a pick-and-place problem will generate a trajectory that works — but it may move joints in sequences, at speeds, and through configurations that look nothing like the human demonstrations used to pretrain a VLA foundation model.

The DATAFARM authors hypothesize, and their results support, that this **behavioral distribution mismatch** is the primary reason raw TAMP data fails to improve VLA fine-tuning. The pretrained model has learned strong priors over motion style and joint-space trajectories from its original dataset. When fine-tuning data comes from a radically different behavioral distribution — planner outputs — the fine-tuning signal fights against those priors rather than refining them.

This is a subtler problem than [sim-to-real transfer](https://humanoidintel.ai/glossary/sim-to-real-transfer) gaps in visual appearance or physics fidelity. It is a *behavioral* domain gap, and it exists even when the simulated physics are perfect.

---

## What DATAFARM Actually Does

DATAFARM intervenes at trajectory generation time, not at training time. Rather than generating TAMP trajectories in an unconstrained way and then trying to filter or reweight them post-hoc, DATAFARM incorporates the pretraining distribution **directly into the TAMP trajectory generation process**. The alignment covers three axes:

- **Robot joint configurations** — ensuring generated poses fall within the distribution of poses seen during pretraining
- **Motion style** — matching velocity profiles, acceleration envelopes, and movement patterns characteristic of the pretraining dataset
- **Temporal execution profiles** — synchronizing the timing structure of generated trajectories with pretraining norms

The paper evaluates across three tabletop manipulation tasks that TAMP can solve directly, plus a cloth-folding task that falls outside TAMP's capability — meaning DATAFARM-augmented fine-tuning must generalize beyond the fine-tuning distribution itself.

---

## The Numbers That Matter

The results reported in the paper are specific enough to carry real weight:

| Condition | Avg. Success Rate |
|---|---|
| Raw TAMP fine-tuning | 8.3% |
| DATAFARM fine-tuning | 56.7% |
| Human teleoperation | 61.7% |

The cloth-folding task — deformable object manipulation outside the fine-tuning distribution — provides a critical generalization probe. The DATAFARM fine-tuned model retains 85% success on this task, compared with 90% for the pretrained baseline. That 5-percentage-point degradation from fine-tuning is notably small: it suggests DATAFARM does not catastrophically overfit the model to the synthetic fine-tuning distribution, which has historically been a concern when adding narrow synthetic data to broad foundation models.

The raw TAMP baseline's 8.3% figure is the headline that should alarm anyone currently planning to dump unconstrained planner output into a VLA fine-tuning pipeline.

---

## What This Means for the Industry

**The data flywheel problem just got a credible partial solution — with important caveats.** Every humanoid platform developing [dexterous manipulation](https://humanoidintel.ai/glossary/dexterous-manipulation) capabilities faces the same fundamental constraint: human teleoperation is expensive, slow, and doesn't scale. TAMP is cheap and infinitely scalable, but — as DATAFARM confirms — naive TAMP data is nearly worthless for VLA fine-tuning.

Companies building proprietary foundation models for humanoid manipulation, including the AI stack players that have raised significant capital on the premise of synthetic data scalability, should take the 8.3% vs. 56.7% gap seriously. It suggests that raw synthetic data pipelines without distribution alignment are likely underperforming in ways that won't be obvious from task-completion metrics alone.

The approach also has implications for how teams structure their pretraining datasets. If the pretraining distribution is well-characterized, DATAFARM's alignment strategy becomes more tractable. If pretraining data is a heterogeneous mixture from multiple robot morphologies and motion styles — a common situation at companies training across fleet variants — the target distribution for alignment becomes ambiguous, and the method's effectiveness may degrade.

**The 85% generalization retention on deformable manipulation is arguably the most important number in the paper.** The fear with any fine-tuning on narrow synthetic distributions is that you trade breadth for precision. DATAFARM appears to navigate that tradeoff more cleanly than prior approaches, though the authors are appropriately cautious: this is tabletop manipulation at research scale, not whole-body dexterous tasks on a 30-DOF humanoid platform.

**The practical deployment path** for this research runs through teams already running TAMP pipelines for motion planning and willing to instrument those pipelines to respect pretraining distribution constraints. That's a non-trivial engineering investment, but substantially cheaper than the alternative: more human teleop hours.

---

## Key Takeaways

- **Raw TAMP trajectories achieve only 8.3% success** when used to fine-tune pretrained VLA models, despite solving target tasks — behavioral distribution mismatch is the identified culprit
- **DATAFARM's distribution-aligned approach reaches 56.7%**, closing most of the gap to human teleoperation (61.7%) without additional human demonstrations
- **Generalization holds**: the fine-tuned model retains 85% success on out-of-distribution deformable object manipulation vs. 90% for the pretrained baseline — minimal catastrophic forgetting
- **The implication for humanoid data pipelines** is significant: synthetic TAMP data generation must be co-designed with the pretraining distribution, not treated as distribution-agnostic
- **Open questions remain** around scalability to multi-DOF humanoid whole-body tasks, heterogeneous pretraining distributions, and real-hardware validation beyond tabletop settings

---

## Frequently Asked Questions

**What is DATAFARM and why does it matter for humanoid robotics?**
DATAFARM is a research framework that aligns task and motion planning (TAMP) generated trajectories with the pretraining distribution of vision-language-action (VLA) models. It matters because synthetic TAMP data is potentially infinitely scalable, but raw TAMP trajectories have been shown to provide minimal benefit for VLA fine-tuning. DATAFARM closes that gap, achieving 56.7% success vs. 8.3% for unaligned TAMP.

**Why does raw TAMP data fail to improve VLA fine-tuning?**
According to the DATAFARM paper, the failure stems from behavioral distribution mismatch: TAMP planners generate trajectories optimized for task completion, not for behavioral consistency with the human demonstration data used to pretrain VLA models. The fine-tuning signal from out-of-distribution trajectories conflicts with the model's learned priors.

**Does DATAFARM fine-tuning hurt generalization to tasks outside the training set?**
The paper reports that the DATAFARM fine-tuned model retains 85% success on deformable object manipulation — a task outside the fine-tuning distribution — compared with 90% for the pretrained baseline. This 5-point degradation is notably small and suggests the method avoids severe catastrophic forgetting.

**How does DATAFARM compare to human teleoperation?**
On the evaluated tabletop manipulation tasks, DATAFARM achieves 56.7% average success versus 61.7% for human teleoperation — a gap small enough to be practically significant, especially given that DATAFARM requires no additional human demonstrations after the initial pretraining phase.

**What are the open limitations of DATAFARM?**
The evaluation is on tabletop manipulation tasks at research scale. Open questions include performance on full humanoid platforms with many degrees of freedom, scalability when the pretraining distribution is heterogeneous across multiple robot morphologies, and real-hardware validation in unstructured environments.