Technical Summary: Contrastive Representation Regularization for Vision-Language-Action Models

Problem Statement

Vision-Language-Action (VLA) models have demonstrated significant potential in robot manipulation by leveraging rich semantic representations from pre-trained Vision-Language Models (VLMs). However, a critical limitation persists: the representations derived from general-purpose VLMs are often suboptimal for robotic control. While VLMs are trained on large-scale visual instruction datasets, they lack explicit exposure to robotic modalities, specifically low-level control actions and proprioceptive information (e.g., joint positions, end-effector states).

Consequently, when VLA models are conditioned on frozen VLM representations, the embeddings tend to be dominated by visual appearance (e.g., scene layout, background textures) rather than control-relevant factors. This misalignment results in representations that fail to capture the robot's physical state or the necessary signals for precise action prediction, leading to suboptimal performance in tasks requiring accurate positioning.

Methodology: Robot State-aware Contrastive Loss (RS-CL)

To address the misalignment between VLM representations and robotic signals, the authors propose Robot State-aware Contrastive Loss (RS-CL). This is a simple, effective representation regularization technique designed to bridge the gap between VLM embeddings and robot proprioceptive states without requiring additional training stages or curated datasets.

Core Components

1. Representation-Level Augmentation (View Cutoff):
   Instead of performing expensive data-level augmentations that require multiple forward passes through the VLM, the authors introduce "view cutoff." This operation randomly masks the feature slice corresponding to a single observation view from the VLM output. This creates alternative embeddings at the representation level with minimal computational overhead, enabling the construction of diverse contrastive pairs efficiently.

2. Learnable Summarization Token:
   To handle the high computational cost of applying contrastive learning on full VLM sequences (which are typically long), a learnable summarization token is appended to the VLM output. The adapter processes this token to produce a compact representative embedding ($z$) used for contrastive learning, while the original embeddings ($h$) remain available for the action decoder.

3. Soft-Weighted Contrastive Objective:
   RS-CL utilizes a supervised contrastive learning framework based on the InfoNCE loss. Unlike standard contrastive losses that treat all negative pairs equally, RS-CL assigns soft weights to similarity scores based on the Euclidean distance between the robot's proprioceptive states ($q_i, q_j$).

   • Mechanism: Pairs of states with small proprioceptive distances (similar physical states) are assigned higher weights, encouraging the model to pull their embeddings closer in the representation space. Pairs with large distances are weighted lower.
   • Formula: The loss is defined as a weighted variant of InfoNCE:
     $$\mathcal{L}_{RS-CL} = - \sum_{i,j} w_{ij} \log \frac{\exp(\text{sim}(z_i, \tilde{z}_j)/\tau)}{\sum_k \exp(\text{sim}(z_i, \tilde{z}_k)/\tau)}$$
     where $w_{ij}$ is computed from the relative distance between proprioceptive states.

4. Training Integration:
   RS-CL operates as an auxiliary objective alongside the standard action prediction loss (e.g., flow-matching loss). The total objective is $\mathcal{L} = \mathcal{L}_{FM} + \lambda \mathcal{L}_{RS-CL}$. The weighting coefficient $\lambda$ is decayed over training, emphasizing representation refinement in early stages and accurate action prediction in later stages. This allows for single-stage, end-to-end training compatible with existing VLA pipelines.

Key Contributions

• Novel Objective: Introduction of RS-CL, which explicitly aligns VLM representations with robot proprioceptive states using continuous distance-based soft supervision.
• Efficient Design: Development of a representation-level augmentation strategy ("view cutoff") that enables lightweight, stable contrastive learning without additional VLM forward passes.
• Compatibility: Demonstration that RS-CL is fully compatible with standard VLA training pipelines, requiring no additional data curation or multi-stage training processes.

Experimental Results

The authors evaluated RS-CL across simulation benchmarks and real-robot experiments.

Simulation Benchmarks

• RoboCasa-Kitchen: RS-CL achieved a state-of-the-art success rate of 69.7% (using 300 demonstrations), surpassing the baseline GR00T N1.5 (65.7%) and other strong baselines like $\pi_0$ (62.5%). Notably, it yielded substantial gains in pick-and-place tasks, improving success rates from 30.3% to 41.5% (+11.2%), a category requiring precise positioning.
• LIBERO: RS-CL improved the average success rate from 95.7% to 96.4%, demonstrating robustness across different multitask suites.
• Efficiency: Compared to Time-Contrastive Networks (TCN), RS-CL achieved higher success rates with significantly lower computational overhead (2.91 $\times 10^{12}$ FLOPs vs. 7.53 $\times 10^{12}$ for TCN), as it avoids temporally structured data mining and extra forward passes.

Real-Robot Experiments

• Setup: Experiments were conducted on a Franka Research 3 arm with a Robotiq 2F-85 gripper.
• Performance: On challenging manipulation tasks, RS-CL improved the success rate from 45.0% to 58.3% (+13.3%).
• Generalization: The method demonstrated improved robustness to visual variations, physical property changes (unseen objects), lighting conditions, and language grounding. It also showed improved performance in tasks with partial occlusion (e.g., closing a lid), attributed to view-invariant representations.

From-Scratch Training

When training VLA models from scratch on various pre-trained VLM backbones (Qwen2.5-VL, SigLIP2, GR00T N1.5), RS-CL consistently improved performance across all architectures. It outperformed baselines that relied on further training VLMs on robotics datasets (e.g., RoboBrain, VeBrain), suggesting that explicit alignment via RS-CL is more effective than general domain adaptation alone.

Significance and Claims

The paper claims that RS-CL effectively addresses the "modality gap" between general-purpose VLMs and the specific demands of robotic control. By explicitly aligning representations with proprioceptive states, the method reshapes the embedding space to capture control-relevant signals (such as task progress and robot pose) rather than just visual appearance.

The authors emphasize that this approach is lightweight and seamless, avoiding the need for complex multi-stage training or large-scale curated robotics datasets. The results suggest that incorporating robot state signals directly into the representation learning process is crucial for accurate action prediction, particularly in tasks requiring precise positioning. The work encourages future research into integrating other sensor signals (e.g., object pose, tactile feedback) to further advance VLA capabilities.