Technical Summary: DISC – Decoupling Instruction from State-Conditioned Control via Policy Generation

1. Problem Statement

The paper addresses a fundamental limitation in current language-conditioned robotic manipulation policies: task-state entanglement. Existing architectures (e.g., Transformers, Diffusion Policies) process natural language instructions and visual observations through shared network parameters. The authors argue this design creates two critical failures:

1. Optimization Tension: Shared parameters must simultaneously learn semantic language understanding (requiring abstraction) and precise visuomotor control (requiring spatial sensitivity), leading to conflicting optimization objectives.
2. Observation Leakage: In limited robotics datasets, visual context often correlates strongly with task identity. Entangled architectures exploit this by learning "scene-to-action" shortcuts, bypassing the language instruction entirely. Consequently, policies succeed on familiar scenes but fail when instructions change or visual ambiguity arises, as they rely on visual priors rather than grounding the linguistic task specification.

2. Methodology: DISC Architecture

The proposed solution, DISC, structurally decouples the task specification (language) from state-conditioned control. Instead of conditioning a universal policy on language, DISC uses a hypernetwork to generate the entire parameter set of a task-specific visuomotor policy based solely on the instruction.

Core Components

• Hypernetwork ($H_\phi$): Takes a language instruction embedding ($e_l$) and outputs the weights ($\theta_\pi$) for a target policy. The target policy ($\pi_{\theta_\pi}$) receives only visual observations ($o_t$) and outputs actions ($a_t$). Because the target policy has no direct access to language, task semantics must be encoded entirely within the generated weights, structurally eliminating observation leakage.
• Two-Stage Generation: Generating coherent high-dimensional weights is challenging. DISC employs a two-stage process to ensure parameter quality:
  1. Weight Initialization Network (WIN): Maps the language embedding to a coarse, semantically informed point in the parameter space (e.g., distinguishing grasping from pushing).
  2. Iterative Refinement Module: Refines the initialization through $T$ learned steps. This module mimics the structure of gradient-based optimization (forward pass, pseudo-gradient estimation, parameter update) but operates as a fully feed-forward neural network. It embeds an inductive bias for optimization without incurring the computational cost of actual gradient descent at inference time.
• Target Policy: A compact, task-specific controller (implemented as a lightweight MLP) configured by the generated weights.

Training and Adaptation

• Training: The system is trained end-to-end via behavior cloning. The loss is backpropagated through the target policy and the hypernetwork to update the hypernetwork parameters ($\phi$).
• Few-Shot Adaptation: For novel tasks, the hypernetwork generates an initialization based on the new instruction. This initialization, already semantically aligned with the task manifold, is then fine-tuned on a small number of demonstrations. This transforms adaptation from a global search into a local refinement, significantly improving sample efficiency.

3. Key Contributions

• Architectural Decoupling: The paper introduces a structural guarantee against observation leakage by ensuring task semantics influence behavior only through generated policy parameters, not through shared processing layers.
• Optimization-Inspired Hypernetwork: A novel hypernetwork design that uses a learned, feed-forward refinement process to generate coherent, high-dimensional policy weights, bridging the gap between expressiveness and control precision.
• Semantic Parameter Manifold: The method learns a parameter space where semantically similar tasks cluster geometrically, enabling robust few-shot adaptation and generalization to paraphrased instructions.

4. Experimental Results

The authors evaluate DISC on simulation benchmarks (LIBERO-90, Meta-World) and a real-world combinatorial benchmark.

• Multi-Task Performance:
  • On LIBERO-90, DISC achieves 94.3% overall success, outperforming the strongest trained-from-scratch entangled baseline (OTTER, 86.6%) by 7.7%.
  • The performance gap widens with task complexity: DISC outperforms baselines by +8.4% on long-horizon tasks.
  • DISC surpasses the large-scale pretrained model $\pi_0$ (91.6%) despite being trained from scratch, and remains competitive with $\pi_{0.5}$ (95.7%).
  • On Meta-World, DISC achieves 92.2% overall, outperforming all baselines.
• Observation Leakage Mitigation (Real-World):
  • In a real-world benchmark where 9 tasks share identical visual contexts (differing only by language), DISC achieves 86.4% success.
  • Entangled baselines (e.g., Octo) drop to 78.5%, often failing to distinguish between tasks with identical visual setups but different instructions, confirming that DISC's decoupling prevents visual shortcuts.
• Few-Shot Adaptation:
  • DISC demonstrates superior sample efficiency. With only 1–3 demonstrations, it achieves non-trivial success rates where diffusion-based methods struggle near 0%.
  • It converges rapidly (within 200 steps) compared to baselines that require longer "warm-up" phases.
• Language Grounding:
  • Under linguistic variation (paraphrased instructions), DISC maintains high performance (85.4% on LIBERO-90), whereas entangled models show significant degradation, indicating DISC grounds task meaning rather than memorizing specific phrasings.

5. Significance and Claims

The paper claims that where language enters a policy architecture matters more than how it is encoded. The authors argue that task-state entanglement is a structural flaw that cannot be fully resolved by scaling data or model size alone.

• Structural Guarantee: By generating task-specific parameters, DISC provides a structural guarantee that task semantics are embedded throughout the controller, preventing the network from learning shortcuts that bypass language.
• Efficiency vs. Pretraining: The results suggest that architectural inductive biases (decoupling) can recover much of the benefit attributed to large-scale pretraining, allowing models trained from scratch to rival or surpass large pretrained Vision-Language-Action (VLA) models.
• Robustness: The decoupled architecture is particularly valuable in complex, long-horizon, or visually ambiguous scenarios where entangled policies fail due to accumulated grounding errors.

The authors conclude that while large pretrained VLAs are complementary for data-rich settings, DISC offers a superior approach for scenarios requiring reliable grounding, efficient control, and adaptation from limited robot data.