# Does Predicting Touch Before Contact Finally Solve Dexterous Manipulation?

A new paper from Zhengyu Tao, Xin Li, and Xin Wang proposes the answer is yes — and the numbers are hard to dismiss. **ForeTac-VLA**, a [vision-language-action model](https://humanoidintel.ai/glossary/vision-language-action-model) augmented with a transformer-based tactile forecasting module, achieves an average success rate of **95% across four real-world contact-rich manipulation tasks**. That is **36.25 percentage points above a fine-tuned VLA baseline** and more than **22 percentage points above state-of-the-art tactile-enhanced VLA methods**, according to the preprint published today on arXiv (2609.20980).

The core insight is deceptively simple: existing tactile-enhanced VLAs are reactive — they process touch signals that have already arrived. ForeTac-VLA instead predicts *future* tactile states before they occur and feeds those predictions into the action generation pipeline alongside observed data. In contact-rich environments — think in-hand re-grasping, compliant assembly, or tasks where the object is partially or fully occluded at the moment of contact — what the robot *expects* to feel is often more actionable than what vision alone can tell it.

The result holds even under degraded visual conditions. The paper reports strong performance maintained under **low-illumination and visually cluttered conditions**, two failure modes that routinely clip the wings of vision-only policies in real deployments.

---

## The Architecture: Forecasting as a First-Class Citizen

ForeTac-VLA's design makes tactile prediction a structural component of the model rather than a post-hoc sensor fusion layer. The pipeline has three distinct stages:

**1. Temporal Tactile Encoding**
Recent tactile observations are encoded into temporal representations — capturing not just the current contact state but its recent history. This is analogous to the role trajectory history plays in whole-body control, giving the model a dynamic context rather than a static snapshot.

**2. Bidirectional Cross-Attention Fusion**
The encoded tactile stream is merged with vision-language features through bidirectional cross-attention. The bidirectional design is notable: tactile context modulates vision-language representations, and vision-language context modulates tactile representations simultaneously. This avoids the hierarchical bottleneck where one modality dominates.

**3. Transformer-Based Multi-Step Forecasting**
A dedicated forecasting module predicts *multi-step future* tactile states. Both observed and anticipated contact representations are then fed into the VLA backbone to condition action generation. The model reasons over a temporal horizon of touch — not just what it feels now, but what it expects to feel across the next several timesteps.

A **ground-truth-to-prediction curriculum** is used during training to handle the period when early forecasts are unreliable. The model is initially supervised on actual future tactile data and gradually transitions to relying on its own predictions — a pragmatic solution to the bootstrapping problem that plagues forecasting-in-the-loop architectures.

---

## Why This Matters for Humanoid [Dexterous Manipulation](https://humanoidintel.ai/glossary/dexterous-manipulation)

The humanoid manipulation stack has a well-documented weak point: [physical AI](https://humanoidintel.ai/glossary/physical-ai) policies trained primarily on visual observations degrade rapidly when contact dynamics are complex or occluded. Picking a soft fruit, threading a connector, or handing an object off between hands all involve contact states that cameras simply cannot resolve at the moment they matter most.

The industry has recognized this. Companies including [Figure AI](https://humanoidintel.ai/companies/figure-ai), [Physical Intelligence (π)](https://humanoidintel.ai/companies/physical-intelligence), and [Skild AI](https://humanoidintel.ai/companies/skild-ai) are building manipulation policies that need to generalize across varied contact conditions. Tactile sensing is increasingly part of hardware roadmaps at leading humanoid developers — but the software stack to exploit it has lagged.

ForeTac-VLA's contribution is methodological: it reframes tactile sensing from a reactive correction signal into a predictive planning input. That framing shift has non-trivial implications. A reactive system corrects after a grasp has already failed; a predictive system adjusts action trajectories *before* the failure mode develops. In assembly tasks or any scenario requiring precise force regulation, the difference between those two approaches can be the difference between 60% and 95% success rates — which is roughly what the paper's numbers suggest.

The 36.25 percentage point gap over a standard fine-tuned VLA is also a credibility marker for the forecasting module specifically. Fine-tuned VLAs already incorporate tactile data as observed input; that comparison is not against a vision-only baseline. Beating it by that margin indicates the *prediction of future contact* — not just the observation of present contact — is doing meaningful work.

---

## Honest Skepticism: What the Paper Doesn't Resolve

A 95% average across four tasks is compelling, but several questions remain unanswered in this preprint:

**Task diversity and generalization.** The paper describes four "real-world contact-rich manipulation tasks" but does not specify what they are in the abstract. Whether those tasks represent the breadth of manipulation challenges a deployed humanoid would face — versus a curated set where tactile forecasting is particularly well-suited — is impossible to assess without the full experimental section.

**Hardware dependency.** ForeTac-VLA's architecture assumes high-quality temporal tactile data as input. The quality of that data is entirely a function of the tactile sensor hardware. Results achieved with a research-grade sensor array may not transfer cleanly to the capacitive or resistive sensors shipping on commercial humanoid hands today, which typically offer lower spatial resolution and higher noise floors.

**[Zero-shot generalization](https://humanoidintel.ai/glossary/zero-shot-generalization).** The paper does not claim zero-shot performance across unseen object categories. Whether the forecasting module generalizes to novel contact geometries — or whether it has implicitly memorized contact dynamics for a fixed task set — is a critical open question for anyone evaluating deployment readiness.

**Latency.** Running a transformer-based forecasting module in the action generation loop adds compute. The paper does not report inference latency figures, which matter enormously for real-time control on embedded hardware.

---

## Industry Trajectory Implications

The broader signal here is that the VLA paradigm is maturing toward multi-modal sensory integration in which vision is one input among several, not the primary one. The field has been pushing in this direction conceptually, but papers demonstrating concrete, reproducible performance gains from non-visual modalities have been relatively sparse.

If ForeTac-VLA's results hold under independent replication — and that's the appropriate standard — they offer a template for how tactile sensing can be elevated from a supplementary feedback signal to a primary planning input. That has direct consequences for robot hand hardware design: sensors that produce clean temporal sequences become more valuable than sensors optimized for high spatial resolution at a single instant.

For humanoid developers currently spec'ing their next-generation end-effectors, the message is clear: the software is now potentially ready to exploit richer tactile data streams than most current hardware provides.

---

## Key Takeaways

- **ForeTac-VLA achieves 95% average success** across four real-world contact-rich manipulation tasks, per the arXiv preprint by Tao, Li, and Wang.
- The model **outperforms a fine-tuned VLA baseline by 36.25 percentage points** and beats state-of-the-art tactile-enhanced VLA methods by over 22 percentage points.
- The core innovation is **predicting future tactile states** via a transformer-based forecasting module, rather than only reacting to observed contact.
- **Bidirectional cross-attention** fuses tactile temporal representations with vision-language features, avoiding single-modality dominance.
- A **ground-truth-to-prediction curriculum** stabilizes training when early tactile forecasts are unreliable.
- Performance holds under **low-illumination and visually cluttered conditions**, addressing a known failure mode of vision-only policies.
- Key open questions: task diversity, hardware-sensor dependency, zero-shot generalization, and inference latency under real-time control constraints.

---

## Frequently Asked Questions

**What is ForeTac-VLA and what problem does it solve?**
ForeTac-VLA is a vision-language-action model that incorporates a transformer-based module to predict future tactile states before they occur. It addresses the limitation of existing VLA models, which rely on visual perception and miss critical contact information in environments where physical interaction states are not visually observable.

**How does ForeTac-VLA compare to existing tactile-enhanced VLA models?**
According to the arXiv preprint, ForeTac-VLA achieves a 95% average success rate across four real-world tasks, outperforming state-of-the-art tactile-enhanced VLA baselines by over 22 percentage points and a fine-tuned VLA baseline by 36.25 percentage points.

**What makes ForeTac-VLA different from a reactive tactile-sensing approach?**
Reactive tactile methods process contact signals after they arrive, correcting errors post-hoc. ForeTac-VLA predicts multi-step future tactile states and uses those predictions to condition action generation before contact occurs, enabling proactive rather than corrective behavior.

**Does ForeTac-VLA work under poor lighting conditions?**
The paper reports that ForeTac-VLA maintains strong performance under low-illumination and visually cluttered conditions — scenarios where vision-only policies typically degrade significantly.

**What are the main limitations to watch before assuming deployment readiness?**
The primary open questions are: the diversity and representativeness of the four evaluation tasks, dependency on high-quality tactile sensor hardware, whether the forecasting module generalizes zero-shot to novel contact geometries, and inference latency under real-time embedded control constraints — none of which are fully resolved in the current preprint.