# Does Embedding 3D World Tokens Inside a VLA Actually Work at Scale?

**98.6% on LIBERO and a real-robot success rate jump from 29.2% to 52.5%** — achieved with zero Gaussian decoding at inference time. That's the headline result from GaussianDream++, a new paper out of a multi-institution team today on arXiv (2608.25659). The work offers a direct answer to one of the more persistent tensions in [Vision-Language-Action model](https://humanoidintel.ai/glossary/vision-language-action-model) design: how do you give a policy rich 3D geometric supervision during training without paying the computational cost of 3D rendering during deployment?

GaussianDream++ inserts what the authors call **World State Tokens** and **World Prediction Tokens** directly into the VLA backbone. A training-only World Representation Head decodes those tokens into current and future scene representations over shared Gaussian primitives. At inference, the head, the renderer, all auxiliary objectives, and the full VGGT/TGE prefix pathway are stripped out entirely. What remains is just 20 world tokens — no online Gaussian decoding, no rollout computation.

On the LIBERO benchmark, the method scores **98.6%**. On the harder LIBERO-Plus variant it scores **87.8%**, with the paper noting particular gains under camera and layout distribution shifts. The real-robot experiments are arguably more meaningful: compared against a reproduced π₀.₅ baseline, GaussianDream++ improves average success from **29.2% to 52.5%** while maintaining closed-loop control efficiency.

---

## The Problem With Existing 3D-Aware VLA Approaches

The VLA manipulation stack has bifurcated into two camps, and both have real liabilities.

Geometry-enhanced policies add 3D structure to anchor the policy to the current scene — useful for grounding, but they don't model how that scene will change as the robot acts. Predictive policies attempt to model future dynamics, but they typically do so in RGB or latent spaces, which can be imprecise about metric 3D structure. More practically, many predictive approaches carry substantial deployment cost: running a world model online at inference adds latency and compute that closes out real-time manipulation.

The original GaussianDream work established that training-time Gaussian reconstruction (current scene) and Gaussian prediction (future scene) provide effective 3D supervision signal. But the paper's authors identify a specific architectural inefficiency: GaussianDream's dense VGGT/TGE-based prefix has to carry state, dynamics, and action-conditioning information simultaneously, which is an expensive and somewhat muddled representational burden.

GaussianDream++ is a targeted fix for that problem, not a wholesale redesign.

---

## The Architecture: Tokens as Geometric Surrogates

The core insight is that you don't need to render Gaussians at inference — you need the policy network to have *internalized* geometric structure during training. The World State Tokens and World Prediction Tokens serve as compact surrogates: during training, the World Representation Head enforces that these tokens carry enough information to decode accurate current and future Gaussian representations. That supervision pressure shapes the token representations. At inference, the tokens retain that geometric information implicitly, without any decoding machinery.

The authors also introduce **static-dynamic factorization**: the model separates persistent scene structure from interaction-relevant motion. This is architecturally sensible — a tabletop, a shelf surface, and a background wall don't need to be re-predicted with every action step. Focusing residual motion prediction on the regions actually affected by manipulation reduces both the modeling burden and, presumably, the noise in the supervision signal.

The 20-token constraint is worth dwelling on. The entire geometric world model compresses to 20 tokens at inference — a remarkably lean footprint for a system making metric 3D predictions during training.

---

## What the Benchmark Numbers Actually Tell You

LIBERO at 98.6% is a strong result, though it's worth noting that LIBERO has become a relatively saturated benchmark — several recent VLA papers now report numbers in the high-90s, which limits its discriminative power. The more informative figure is LIBERO-Plus at 87.8%, where distribution shift (camera viewpoint changes, layout rearrangements) is the evaluation pressure. The real-world claim of 52.5% versus a 29.2% baseline is striking, though baseline reproduction methodology matters enormously for real-robot comparisons and the paper would need detailed scrutiny on that front.

The gains under camera and layout shifts are directionally consistent with the core architectural claim: if the policy has genuinely internalized metric 3D structure, it should be more robust to changes in viewpoint and scene configuration than a policy that only sees RGB.

---

## Industry Implications for the VLA Stack

For teams building [dexterous manipulation](https://humanoidintel.ai/glossary/dexterous-manipulation) policies on humanoid platforms, GaussianDream++ represents a meaningful existence proof: **3D world modeling can be a training-time regularizer rather than a deployment-time cost center**. The practical constraint for teams like [Physical Intelligence (π)](https://humanoidintel.ai/companies/physical-intelligence) or [Skild AI](https://humanoidintel.ai/companies/skild-ai) isn't benchmark accuracy in controlled settings — it's inference latency, hardware memory footprint, and robustness to real-world distribution shift. A method that delivers geometric supervision benefits while adding only 20 tokens to the inference graph is architecturally compatible with real deployment constraints.

The static-dynamic factorization approach also points toward an underexplored dimension of world model design for manipulation: most of a manipulation scene is static most of the time. Policies that model everything with equal resolution are wasting capacity.

The key open question is generalization beyond tabletop manipulation benchmarks. Does the 3D supervision signal remain useful when the robot's morphology changes (different end-effector geometries, different camera placements), or when the task involves contact dynamics complex enough that Gaussian primitives are a poor scene representation? The paper doesn't answer that, and it's the right question to ask before integrating this approach into a broader humanoid manipulation stack.

---

## Key Takeaways

- **GaussianDream++ achieves 98.6% on LIBERO and 87.8% on LIBERO-Plus**, with clear robustness gains under camera and layout distribution shifts.
- **Real-robot success improves from 29.2% to 52.5%** over a reproduced π₀.₅ baseline, while maintaining closed-loop efficiency.
- **The method uses only 20 world tokens at inference** — all Gaussian decoding, rendering, and auxiliary objectives are training-only; the inference graph stays lean.
- **Static-dynamic factorization** separates persistent scene structure from interaction-relevant motion, focusing modeling capacity where it matters.
- **The core argument**: 3D world modeling should be a training-time supervision signal, not a deployment-time compute cost — a framing with direct implications for real-hardware VLA deployment.

---

## Frequently Asked Questions

**What is GaussianDream++?**
GaussianDream++ is a VLA policy architecture that uses World State Tokens and World Prediction Tokens to inject 3D Gaussian-based geometric supervision during training. At inference, all decoding machinery is removed, leaving only 20 compact tokens. It scores 98.6% on the LIBERO manipulation benchmark and improves real-robot success from 29.2% to 52.5% compared to a π₀.₅ baseline.

**How does GaussianDream++ differ from the original GaussianDream?**
The original GaussianDream uses a dense VGGT/TGE-based prefix that carries state, dynamics, and action-conditioning simultaneously. GaussianDream++ replaces this with a compact, policy-native token insertion approach, eliminating online Gaussian decoding at inference and significantly reducing deployment cost.

**What is LIBERO-Plus and why does it matter?**
LIBERO-Plus is a harder variant of the LIBERO manipulation benchmark that introduces distribution shifts including camera viewpoint changes and layout rearrangements. High performance on LIBERO-Plus is a stronger indicator of generalization than the standard LIBERO suite, which has become relatively saturated in recent VLA research.

**What does "static-dynamic factorization" mean in this context?**
It means the model separately represents persistent scene structure (surfaces, objects not being manipulated) and dynamic, interaction-relevant regions. This allows the world model to focus its predictive capacity on the parts of the scene actually affected by robot actions, rather than modeling everything at equal resolution.

**Is GaussianDream++ ready for deployment on full humanoid platforms?**
The paper demonstrates results on tabletop manipulation tasks. Whether the 3D Gaussian supervision signal generalizes to full humanoid morphologies — with different end-effector geometries, moving camera platforms, and complex whole-body contact dynamics — is an open question the current work does not address.