Research map / Architectures and efficiency / RNNs and normalization
Distilling Morphology-Conditioned Hypernetworks for Efficient Universal Morphology Control
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Distills a universal transformer teacher into a morphology-conditioned hypernetwork that generates a per-robot MLP policy, matching teacher performance on UNIMAL with 6-14x smaller models and 67-160x fewer FLOPs.
Problem
Universal morphology control needs transformer policies for strong zero-shot performance across robot morphologies, but their memory and computation cost at inference is much higher than MLPs.
Approach
A morphology-conditioned hypernetwork generates the weights of a small MLP policy once per robot, so inference runs like an MLP; the hypernetwork is trained via behavior-cloning distillation from a universal transformer teacher, with choices (teacher type, architecture alignment, number of distillation robots, task-space dropout) tuned for generalization to unseen morphologies.
Key results
- HyperDistill matches the ModuMorph teacher on training and unseen test robots in FT, VT and Obstacle environments on UNIMAL.
- HyperDistill reduces model size by 6-14x and inference FLOPs by 67-160x relative to ModuMorph (oracle).
- Dropout on the context embedding improves HyperDistill generalization by 8.5% in FT, and increasing PD robots from 100 to 1,000 raises HyperDistill generalization by 6%, 15%, 13% across the three environments.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HyperDistill instead of the universal ModuMorph teacher | Universal morphology control · UNIMAL training and test robots across FT, VT, and Obstacle | performance | No clear effect |
| Use HyperDistill instead of the universal ModuMorph teacher | Universal morphology control · FT, VT, and Obstacle | inference model size: 6-14 times | Helped |
| Use HyperDistill instead of the universal ModuMorph teacher | Universal morphology control · FT, VT, and Obstacle | inference FLOPs: 67-160 times | Helped |
| Use a compressed standard TF student | Universal morphology control · FT, VT, and Obstacle | performance | Hurt |
| Use compressed ModuMorph rather than compressed standard TF | Universal morphology control · FT, VT, and Obstacle | performance | Helped |
| Use compressed ModuMorph rather than HyperDistill | Universal morphology control · FT, VT, and Obstacle | generalization performance | Hurt |
| Use multi-robot MLP rather than HyperDistill | Universal morphology control · UNIMAL training and test robots | performance | Hurt |
| Train the HN policy with RL rather than policy distillation | HN policy · UNIMAL training and test robots | performance | Hurt |
| Distill from single-robot MLP teachers instead of a universal TF teacher (ablation) | HyperDistill and TF (oracle) students · FT training and test robots | generalization performance | Hurt |
| Use a TF student aligned with the universal TF teacher instead of HyperDistill (ablation) | Student distilled from a universal TF teacher · FT test robots | generalization performance | Helped |
| Use HyperDistill rather than TF student with single-robot MLP teachers (ablation) | Student distilled from single-robot MLP teachers · FT test robots | generalization performance | Helped |
| Increase PD robots from 100 to 1,000 (ablation) | HyperDistill · FT, VT, and Obstacle test robots | generalization performance: 6%, 15% and 13% | Helped |
Surprising findings
- Distilling from single-robot MLP teachers generalizes much worse than distilling from a universal transformer teacher, despite similar training-robot performance.
- Training the hypernetwork policy directly via RL (HN-RL) performs poorly on both training and test robots.
Limitations
- Hypernetwork must be called whenever task context changes, which may reduce efficiency gains in settings with language-conditioned or time-varying tasks.
- Evaluated only on the UNIMAL benchmark with locomotion tasks in three environments.
- Generalization gap from teacher-student architecture misalignment remains unavoidable.
Experimental setup
Datasets: UNIMAL
Benchmarks: UNIMAL
Metrics: episodic return on training and test robots model size (parameters) FLOPs at inference
Baselines: ModuMorph (oracle) TF (compressed) ModuMorph (compressed) Multi-robot MLP HN-RL
Builds on: ModuMorph UNIMAL hypernetworks policy distillation
Closest papers by meaning
- DittoGym: Learning to Control Soft Shape-Shifting Robots (ICLR 2024 · similarity 0.87)
- Convergent Functions, Divergent Forms (NeurIPS 2025 · similarity 0.86)
- Scaling Proprioceptive-Visual Learning with Heterogeneous Pre-trained Transformers (NeurIPS 2024 · similarity 0.85)
- $\textit{Hyper-GoalNet}$: Goal-Conditioned Manipulation Policy Learning with HyperNetworks (NeurIPS 2025 · similarity 0.84)
- UniGraspTransformer: Simplified Policy Distillation for Scalable Dexterous Robotic Grasping (CVPR 2025 · similarity 0.84)
- Zero-Shot Robotic Manipulation with Pre-Trained Image-Editing Diffusion Models (ICLR 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.