DISC: Decoupling Instruction from State-Conditioned Control via Policy Generation
📰 ArXiv cs.AI
Learn how DISC decouples instruction from state-conditioned control via policy generation to improve language-conditioned manipulation policies
Action Steps
- Implement a hypernetwork to generate policy parameters
- Decouple instruction from state-conditioned control using DISC
- Train a language-conditioned manipulation policy using the generated parameters
- Evaluate the policy using metrics such as success rate and observation leakage
- Compare the performance of DISC with traditional approaches
Who Needs to Know This
AI researchers and engineers working on language-conditioned manipulation policies can benefit from this approach to improve policy generation and reduce observation leakage
Key Insight
💡 Decoupling instruction from state-conditioned control can reduce observation leakage and improve language grounding in manipulation policies
Share This
🤖 Improve language-conditioned manipulation policies with DISC, a new approach that decouples instruction from state-conditioned control via policy generation
Key Takeaways
Learn how DISC decouples instruction from state-conditioned control via policy generation to improve language-conditioned manipulation policies
Full Article
Title: DISC: Decoupling Instruction from State-Conditioned Control via Policy Generation
Abstract:
arXiv:2605.20856v1 Announce Type: cross Abstract: Language-conditioned manipulation policies typically process instructions and observations through shared network parameters. This task-state entanglement provides a pathway for observation leakage -- networks learn scene-to-action shortcuts that bypass language grounding entirely. DISC eliminates this failure structurally. Rather than conditioning a universal policy on language, DISC uses a hypernetwork to generate the entire parameter set of a
Abstract:
arXiv:2605.20856v1 Announce Type: cross Abstract: Language-conditioned manipulation policies typically process instructions and observations through shared network parameters. This task-state entanglement provides a pathway for observation leakage -- networks learn scene-to-action shortcuts that bypass language grounding entirely. DISC eliminates this failure structurally. Rather than conditioning a universal policy on language, DISC uses a hypernetwork to generate the entire parameter set of a
DeepCamp AI