Факультет інформатики

Permanent URI for this community

Now showing 1 - 1 of 1

Efficient Policy Learning via Knowledge Distillation for Robotic Manipulation
(2026) Severhin, Oleksandr ; Kriukova, Halyna
This research analyzes computational bottlenecks and potential fixes for deploying high-capacityWorld Models on resource-constrained edge devices. The TDMPC- Opt framework [1] provides a fix: compressing a massive 317-million-parameter teacher model trained natively on MT30 into a lightweight 1-million-parameter student model that can be further quantized and deployed. This study replaces a static distillation coefficient with mathematically formulated dynamic distillation coefficients, such as Linear Decay, Cosine Decay, and Four Phase, to optimize the student’s learning process. Furthermore, to improve model learning, a sparse Mixtureof- Experts (MoE) architecture with Top-K routing was implemented. The models were evaluated across both the Meta-World (MT30) benchmark and the modern, complex ManiSkill3 physics environment. Results demonstrate that dynamic schedules, such as Cosine Decay, boost dense reward accumulation by up to 40.5% in complex spatial tasks compared to static baselines.