Technical Summary: Generalizable VLA Finetuning via Representation Anchoring and Language-Action Alignment

1. Problem Statement

Vision-Language-Action (VLA) models have emerged as a standard approach for robot manipulation, typically created by finetuning a pretrained Vision-Language Model (VLM) on robot demonstrations via Behavior Cloning (BC). While this approach transfers the VLM's semantic priors (understanding of spatial layout, color, shape, etc.) to control policies, standard BC finetuning suffers from two critical failure modes that degrade generalization:

1. Catastrophic Forgetting of Representations: Standard BC optimizes only the action prediction loss. Over the course of finetuning, the updates progressively overwrite the pretrained visuolinguistic and spatial concepts acquired during internet-scale pretraining. Consequently, a policy trained to pick up a "green mug" may fail to generalize to a "pink mug" in the same scene, or fail to adapt when object positions are swapped, because the backbone has lost its color and spatial reasoning capabilities.
2. Language-Action Misalignment: Common remedies, such as co-training on general web image-text data (e.g., VQA, captioning), apply language and action losses to separate observations. This leaves the VLA with disjoint supervision: the action head learns from expert demonstrations while the language head learns from generic tasks. Because they are never supervised on the same robot observation, their predictions can contradict each other (e.g., the action head moves "right" while the language head predicts "left"), a misalignment that standard benchmarks often fail to expose.

2. Methodology: Anchor-Align

The authors propose Anchor-Align, a finetuning recipe that augments standard BC with two specific objectives to preserve pretrained representations and align language with action. The method requires no additional data, annotation, or architectural changes beyond the standard VLA setup.

2.1 Base Architecture

The method operates on VLA architectures that use a pretrained VLM backbone with a dedicated action head (supporting both regression and flow-matching action heads). The backbone encodes visual observations and language instructions, while a state projector encodes robot proprioception. The action head predicts continuous action chunks based on these multimodal representations.

2.2 Vision-Language Anchoring

To prevent catastrophic forgetting, Anchor-Align introduces a distillation loss that anchors the trainable VLA backbone to a frozen copy of the original pretrained VLM (the "Anchor VLM").

• Mechanism: The frozen Anchor VLM processes the same input batch (images and text) in parallel. At every decoder layer $u$, the hidden states of the vision and text tokens in the trainable backbone ($H^S_u$) are forced to match those of the frozen anchor ($H^A_u$).
• Loss Function: The anchoring loss is the squared Frobenius norm of the difference between these hidden states, averaged over all decoder layers:
  $$L_{anchor} = \frac{1}{|D|} \sum_{u \in D} \| H^S_u[m] - H^A_u[m] \|_F^2$$
  where $m$ denotes the set of vision and text token positions. This acts as a regularizer in representation space, preserving the semantic substrate without requiring weight-space constraints.

2.3 Language-Action Alignment

To resolve the misalignment between language and action, the method introduces a joint supervision objective where both heads are trained on the same robot observation.

• Mechanism: The continuous action targets from expert demonstrations are programmatically converted into discrete motion-direction labels (e.g., "up," "down," "left," "right," "forward," "backward"). This is achieved by averaging the translational components of action chunks, filtering near-stationary samples, and discretizing the dominant axis.
• Loss Function: The model is trained to predict this discrete direction label using the language head's output on the same observation used for action prediction. Specifically, the last hidden state of the instruction token (which summarizes the visual prefix and instruction) is projected through the frozen pretrained language head to generate logits over the vocabulary. The alignment loss is the cross-entropy between these logits and the ground-truth direction label:
  $$L_{align} = CE(\hat{a}_{lang}, a_{lang})$$
• Total Objective: The final training loss combines the standard BC action loss ($L_{action}$) with the anchoring and alignment losses:
  $$L_{total} = L_{action} + \lambda_{anchor} L_{anchor} + \lambda_{align} L_{align}$$

3. Key Contributions

1. Identification of Dual Failure Modes: The paper identifies that standard BC causes representation drift (forgetting) and that co-training fails to align language and action on the same observation, leading to contradictory predictions.
2. Anchor-Align Framework: A novel finetuning recipe that combines Vision-Language Anchoring (to preserve pretrained representations via distillation) and Language-Action Alignment (to ground language predictions in action targets on the same observation).
3. Data-Efficient Solution: The method achieves significant improvements without requiring new datasets, human annotations, or architectural modifications, relying solely on the existing demonstration data and the frozen VLM.
4. Diagnostic Tool: The alignment framework provides a direct metric for diagnosing language-action misalignment in co-trained VLAs, showing that better alignment correlates with higher task success.

4. Experimental Results

4.1 Simulation Benchmarks

The method was evaluated on LIBERO-PRO, LIBERO-Plus, and CALVIN benchmarks using two distinct VLA architectures (VLA-Adapter and StarVLA).

• LIBERO-PRO (OOD Generalization): Anchor-Align significantly outperformed all baselines, including co-trained models. Notably, on the "Position Swap" test (where object locations are rearranged), standard BC and co-trained baselines scored near 0% (replaying memorized trajectories), while Anchor-Align achieved 22.6%.
• LIBERO-Plus (Perceptual Robustness): Anchor-Align improved performance across all seven perturbation axes (camera view, lighting, background, etc.), with the largest gains on background texture (+8.9%) and sensor noise (+7.4%).
• CALVIN (Long-Horizon): On the long-horizon ABC→D benchmark, Anchor-Align achieved the highest success rate for completing chains of five instructions (77.9% vs. 66.5% for OpenVLA-OFT), demonstrating that preserving semantics improves compounding error resistance.

4.2 Real-World Experiments

Experiments were conducted on a physical xArm7 robot under challenging conditions including spatial rearrangements, cluttered scenes, and semantic perturbations.

• Generalization: Anchor-Align improved real-robot success rates from 28% to 54% (VLA-Adapter) and 37% to 60% (StarVLA) across unseen spatial and semantic perturbations.
• Semantic Perturbation: In a critical test where a policy trained on a "green mug" was instructed to pick up a "pink mug," standard BC failed 90% of the time (defaulting to the green mug), whereas Anchor-Align succeeded 100% of the time in picking the correct object.
• Error Analysis: Failure mode analysis showed Anchor-Align nearly eliminated "Semantic Errors" (7 to 0) and "Incorrect Object" approaches (10 to 1) compared to standard BC.

4.3 Ablation and Analysis

• Necessity of Both Terms: Ablation studies confirmed that both Anchoring and Alignment are necessary; using either alone improved performance over standard BC, but combining them yielded the best results.
• Alignment vs. Regularization: Control experiments with corrupted labels (Shuffle and Scatter) showed that the gains are not merely due to regularization effects but stem from genuine language-action alignment.
• Correlation: A strong positive Pearson correlation (+0.51) was found between the degree of language-action alignment and task success, whereas co-trained baselines showed near-zero correlation.
• Representation Preservation: Visual reasoning tests (GQA) showed that while standard BC lost 94% of its reasoning accuracy during finetuning, Anchor-Align retained 70%.

5. Significance and Claims

The paper claims that preserving pretrained representations and effective action learning are not fundamentally at odds. By recasting VLA finetuning as a process of preserving and aligning pretrained VLM priors rather than trading them away for control, Anchor-Align offers a simple, data-efficient recipe for adapting foundation models to embodied tasks.

The authors emphasize that their method:

• Solves the "catastrophic forgetting" problem inherent in standard BC.
• Fixes the "language-action misalignment" problem inherent in co-training.
• Demonstrates that these improvements hold across different architectures (regression and flow-matching) and scales (0.5B to 3B parameters).
• Provides a path toward world models that retain broad spatiotemporal priors while remaining controllable and semantically grounded.

The work suggests that future VLA development should focus on anchoring representations and aligning modalities on the same observation, rather than simply scaling up data or model size.