DISC: Decoupling Instruction from State-Conditioned Control via Policy Generation
The paper titled 'DISC: Decoupling Instruction from State-Conditioned Control via Policy Generation' introduces a new method for generating visuomotor policies without direct access to language. This approach aims to eliminate observation leakage and improve task-awareness by using a hypernetwork to create task-specific policies from instructions alone. The results show that DISC outperforms existing methods, particularly in complex tasks, and demonstrates robust generalization capabilities.
- ▪DISC uses a hypernetwork to generate task-specific visuomotor policies from language instructions without direct access to the language.
- ▪The method addresses the issue of observation leakage that occurs in traditional language-conditioned manipulation policies.
- ▪Experimental results indicate that DISC outperforms all entangled baselines on various benchmarks, especially in complex, long-horizon tasks.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Robotics arXiv:2605.20856 (cs) [Submitted on 20 May 2026] Title:DISC: Decoupling Instruction from State-Conditioned Control via Policy Generation Authors:Hanxiang Ren, Pei Zhou, Xunzhe Zhou, Yanchao Yang View a PDF of the paper titled DISC: Decoupling Instruction from State-Conditioned Control via Policy Generation, by Hanxiang Ren and 2 other authors View PDF HTML (experimental) 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.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
