Source note
Breaking Lock-In: Preserving Steerability under Low-Data VLA Post-Training
Summary
DeLock targets a common failure in low-data post-training of vision-language-action policies: after fine-tuning on a narrow demo set, the robot learns the task but stops following new instructions. It preserves the model's pre-trained visual grounding during fine-tuning and adds a test-time prompt-contrast rule that steers action generation toward a new instruction.
Problem
- Low-data supervised fine-tuning can cause lock-in: the policy overfits to the post-training demos and ignores novel instructions even when it knows the underlying concepts from pretraining.
- The paper splits this into concept lock-in (fixation on trained objects or attributes) and spatial lock-in (fixation on trained locations or relations such as left/right or upper/lower).
- This matters because collecting broad robot demonstration coverage is expensive, so practical adaptation often happens with only 80-100 demos per task and narrow instruction coverage.
Approach
- Visual encoder drift regularization: during post-training, DeLock adds an L2 penalty on the visual encoder weights relative to the pre-trained model, with objective
L_BC + λ||θ_v - θ_v^pre||^2, to keep visual grounding from collapsing onto the narrow fine-tuning distribution. - Contrastive Prompt Guidance (CPG): at test time, the same flow-based policy is run with a novel prompt
τ+and a trained promptτ-, then their denoising vector fields are combined asv_CPG = v(τ-) + w(v(τ+) - v(τ-))to push action generation toward the novel instruction. - The trained prompt acts as a negative condition because it captures the post-training bias; the novel prompt is the positive condition.
- The method does not rely on extra supervision, external foundation-model labels, auxiliary losses, or augmented datasets.
- The evaluation suite includes 8 tasks: 4 LIBERO simulation tasks with 100 demos each and 4 DROID real-world tasks with 80 demos each, designed to test concept and spatial lock-in directly.
Results
- On the 8-task evaluation with 20 trials per task, DeLock reports strong performance across OOD-location and novel-instruction tests: T1 16/20, T4 OOD 15/20, T2[C] 19/20, T3[C] 19/20, T4[C] 17/20, T5[S] 11/20, T6[S] 13/20, T7[S] 14/20, T8[C+S] 13/20.
- Against the low-data baseline RETAIN, DeLock is better on every novel-prompt task shown in Table 2. Examples: T2 19/20 vs 0/20, T5 11/20 vs 0/20, T7 14/20 vs 2/20, T8 13/20 vs 1/20.
- Against the high-resource reference π0.5-DROID, DeLock matches or exceeds it on several reported real-world novel-instruction tasks: T4[C] 17/20 vs 18/20, T7[S] 14/20 vs 11/20, T8[C+S] 13/20 vs 0/20. The paper frames this as reaching or beating a state-of-the-art generalist policy that used much more curated post-training data.
- Ablations show both parts matter. Removing visual regularization drops performance sharply, for example T2 19/20 → 9/20 and T8 13/20 → 0/20. Removing CPG keeps some concept performance but fails on spatial lock-in tasks: T5/T6/T7/T8 = 0/20, 0/20, 0/20, 0/20 without CPG, versus 11/20, 13/20, 14/20, 13/20 with full DeLock.
- Freezing the visual encoder is weaker than regularized adaptation: Frozen-Vis gets T5 2/20, T6 11/20, T7 8/20, T8 4/20, below full DeLock on the same tasks.
- The paper also gives qualitative evidence that standard fine-tuning collapses vision-language attention onto trained targets, while DeLock shifts attention and denoising trajectories in line with the new prompt.