DeLock Framework: Mitigating VLA Lock-in

• The DeLock Framework is a novel approach that prevents lock-in in VLA policy fine-tuning by combining weight drift regularization with contrastive prompt guidance.
• It preserves prompt-conditioned visual grounding by applying an ℓ2 penalty on the visual encoder weights, maintaining consistent object and spatial representations.
• The framework outperforms baselines by achieving high success rates in both novel concept and spatial tasks, demonstrating robust generalization in simulation and real-world scenarios.

The DeLock framework addresses the phenomenon of “lock-in” in low-data post-training of generalist vision-language-action (VLA) policies, whereby supervised fine-tuning (SFT) on a small demonstration dataset causes loss of prompt-conditioned generalization. DeLock employs a two-fold strategy—preserving the visual grounding by regularizing weight drift in the visual encoder and steering denoising dynamics at inference via contrastive prompt guidance. This design mitigates both concept and spatial lock-in, enabling robust generation under novel instructions without reliance on external supervision or large-scale data augmentation (Huang et al., 25 Apr 2026).

1. The Lock-In Phenomenon in VLA Policy Post-Training

Lock-in occurs during low-data SFT of a VLA policy, where the resultant fine-tuned model $\pi_{\theta_{\mathrm{ft}}}(a|o, \tau)$ becomes over-specialized to post-training instruction-support $\mathcal{T}_\star$ and ceases to respond appropriately to novel prompts. Two primary failure modes are identified:

• Concept lock-in: The policy fixates on objects or attributes present in the training data. Formally, $$\pi_{\theta_{\mathrm{ft}}}(a|o, \tau^+) \approx \pi_{\theta_{\mathrm{ft}}}(a|o, \tau^-)$$, even when $\tau^+ \ne \tau^-$, reflecting insensitivity to changes in instruction.
• Spatial lock-in: The model remains bound to post-trained spatial targets (e.g., always acting on the “lower microwave” regardless of novel prompts).

This insensitivity emerges despite the policy retaining the relevant skill; the fine-tuning process degrades responsiveness to out-of-distribution instructions while preserving in-distribution (trained prompt) competence (Huang et al., 25 Apr 2026).

2. Preserving Visual Grounding: Weight-Drift Regularization

VLA policies are modular, typically comprising a visual encoder $v_{\theta_v}(o)$, a language backbone $l_{\theta_\ell}(\tau, \cdot)$, and an action expert $a_{\theta_a}(\cdot)$. During standard SFT—even with parameter-efficient schemes such as LoRA adapters—unconstrained drift in the visual encoder $\theta_v$ erodes the grounding of object and spatial semantics.

DeLock addresses this by introducing an $\ell_2$ penalty on the distance between the fine-tuned and pre-trained visual encoder weights:

$$\mathcal{L}_{\mathrm{DeLock}}(\theta) = \mathcal{L}_{\mathrm{BC}}(\theta; \mathcal{D}_\star) + \lambda \lVert \theta_v - \theta_v^{\mathrm{pre}} \rVert_2^2$$

where $\mathcal{T}_\star$0 controls the regularization strength. Only LoRA adapters in the language and action modules are updated without regularization, while $\mathcal{T}_\star$1 is free to adapt under the penalty, preserving prompt-conditioned visual grounding essential for generalization (Huang et al., 25 Apr 2026).

3. Test-Time Contrastive Prompt Guidance (CPG)

DeLock employs CPG at inference to steer policy rollouts toward novel target prompts by leveraging the structure of denoising diffusion over the action space. At each diffusion-step $\mathcal{T}_\star$2 and rollout-step $\mathcal{T}_\star$3:

• The policy predicts denoising vector fields $\mathcal{T}_\star$4 (for the trained prompt) and $\mathcal{T}_\star$5 (for the novel prompt).
• The CPG vector is constructed:

$\mathcal{T}_\star$6

where $\mathcal{T}_\star$7 is a guidance scale hyperparameter.

• Action update: $\mathcal{T}_\star$8

This directs denoising toward dynamics consistent with the novel prompt while suppressing the default bias toward trained behaviors. The process can be interpreted as imposing a test-time contrastive bias that amplifies prompt-unique information encoded in the denoising field (Huang et al., 25 Apr 2026).

4. Algorithms and Implementation

DeLock consists of two algorithmic components, both detailed in the primary reference:

• Algorithm 1 (Low-Data SFT with Visual Encoder Regularization): Initializes with frozen pre-trained weights for the visual encoder; LoRA adapters are deployed on language/action modules. The SFT loop simultaneously minimizes behavioral cloning loss and the $\mathcal{T}_\star$9 regularization term, updating $$\pi_{\theta_{\mathrm{ft}}}(a|o, \tau^+) \approx \pi_{\theta_{\mathrm{ft}}}(a|o, \tau^-)$$0 and LoRA parameters accordingly.
• Algorithm 2 (Test-Time CPG): At each inference step, actions are iteratively refined using the contrastive update rule. — For a complete acquisition, two prompt encodings must be available: the base trained prompt $$\pi_{\theta_{\mathrm{ft}}}(a|o, \tau^+) \approx \pi_{\theta_{\mathrm{ft}}}(a|o, \tau^-)$$1 and the novel prompt $$\pi_{\theta_{\mathrm{ft}}}(a|o, \tau^+) \approx \pi_{\theta_{\mathrm{ft}}}(a|o, \tau^-)$$2.

Training employs AdamW optimizer, batch size 32, cosine learning rate schedule, and regularization parameter $$\pi_{\theta_{\mathrm{ft}}}(a|o, \tau^+) \approx \pi_{\theta_{\mathrm{ft}}}(a|o, \tau^-)$$3 typically in $$\pi_{\theta_{\mathrm{ft}}}(a|o, \tau^+) \approx \pi_{\theta_{\mathrm{ft}}}(a|o, \tau^-)$$4. Inference guidance scale $$\pi_{\theta_{\mathrm{ft}}}(a|o, \tau^+) \approx \pi_{\theta_{\mathrm{ft}}}(a|o, \tau^-)$$5 is selected from $$\pi_{\theta_{\mathrm{ft}}}(a|o, \tau^+) \approx \pi_{\theta_{\mathrm{ft}}}(a|o, \tau^-)$$6, and diffusion steps are set to ensure 20–50 Euler steps per trajectory (Huang et al., 25 Apr 2026).

Component	Role	Regularization
Visual	$$\pi_{\theta_{\mathrm{ft}}}(a|o, \tau^+) \approx \pi_{\theta_{\mathrm{ft}}}(a|o, \tau^-)$$7	$$\pi_{\theta_{\mathrm{ft}}}(a|o, \tau^+) \approx \pi_{\theta_{\mathrm{ft}}}(a|o, \tau^-)$$8 drift (Eq. 1)
Language	$$\pi_{\theta_{\mathrm{ft}}}(a|o, \tau^+) \approx \pi_{\theta_{\mathrm{ft}}}(a|o, \tau^-)$$9, LoRA	None
Action	$\tau^+ \ne \tau^-$0, LoRA	None

5. Evaluation Protocols and Quantitative Results

DeLock is evaluated on an 8-task benchmark suite covering both simulation (LIBERO-based) and real-world manipulation (DROID), with 100/80 demonstrations per task in simulation/real world. Success metrics are defined as number of successful completions in 20 held-out trials under in-distribution (trained prompt) and out-of-distribution (novel prompt or location) scenarios. Baselines include RETAIN, a high-resource generalist (π_{0.5}-DROID), and spatial-forcing methods relying on additional supervision.

Main findings:

• DeLock achieves 17–19/20 success on concept-probe (novel object/attribute) tasks and 11–14/20 on spatial-probe (novel spatial target) tasks, surpassing all baselines.
• RETAIN exhibits 0–6/20 on novel prompts, and even π_{0.5}-DROID (high-resource generalist) fails on fine-grained spatial generalization (e.g., T8=0/20).
• Ablation confirms necessity of both visual regularization and CPG: without them, generalization collapses, producing either training-biased or spatially-fixed outputs (Huang et al., 25 Apr 2026).

6. Analysis, Qualitative Findings, and Limitations

Attention-visualization demonstrates that DeLock maintains visual-language correspondence, shifting focus according to prompt content, in contrast to standard SFT policies, which lose this alignment. Denoising-dynamics analysis reveals that CPG vectors consistently alter action trajectories toward prompt-implied spatial or conceptual targets. Real-world robotic rollouts validate these findings, with DeLock reliably executing novel-instruction tasks while baseline policies revert to post-training biases.

A documented failure case occurs with malformed prompts; here, CPG essentially reverts to the trained bias, suggesting the necessity of valid semantic content in $\tau^+ \ne \tau^-$1 for successful generalization.

Significance lies in showing that neither external foundation-model supervision nor large data quantities are essential for recoverable generalization. DeLock leverages pre-trained grounding and denoising dynamics for instruction-sensitive behavior under extreme data paucity, surpassing even high-resource alternatives on OOD generalization (Huang et al., 25 Apr 2026).