Research map / Architectures and efficiency / Pruning and efficient tuning
MetaMorph: Learning Universal Controllers with Transformers
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Condition the Transformer controller on morphology information (ablation) | MetaMorph · Flat terrain; dynamics and kinematics variations across all environments | mean reward | Helped |
| Use learned position embeddings instead of omitting positional embeddings (ablation) | MetaMorph · Flat terrain | mean reward | Helped |
| Use learned position embeddings instead of hand-designed limb IDs (ablation) | MetaMorph · Flat terrain | mean reward | Helped |
| Add hand-designed unique limb IDs (ablation) | MetaMorph-NPE · Flat terrain | mean reward | Helped |
| Add morphology information and joint angle ranges to Amorpheus-style observations (ablation) | MetaMorph-AO · Flat terrain | mean reward | Helped |
| Add dynamic replay buffer balancing (ablation) | MetaMorph · Flat terrain, Variable terrain, Obstacles | robots reaching 75% of MLP baseline performance: 10-15% | Helped |
| Replace Transformer processing with GNN processing | Jointly trained robot controller · Flat terrain, Variable terrain, Obstacles | mean reward: 2 to 3 times below MetaMorph | Hurt |
| Pre-train on a collection of morphologies before fine-tuning on new morphologies | MetaMorph · Unseen morphologies, same task | training iterations for comparable reward: 2 to 3 times | Helped |
| Pre-train on locomotion tasks before fine-tuning on new morphologies and novel tasks | MetaMorph · Escape; Obstacles (cylinders) | training iterations for comparable reward: 2 to 3 times | Helped |
| Train with a fixed depth-first node ordering rather than randomizing visitation order among nodes at the same depth (ablation) | MetaMorph · Zero-shot transfer across robot morphologies | performance: ∼ 75% | Hurt |
| Use joint angle ranges, module geometry, material, and other morphology descriptors as conditioning inputs | MetaMorph · Unseen dynamics and kinematics variations | zero-shot reward | Helped |
Closest papers by meaning
- Distilling Morphology-Conditioned Hypernetworks for Efficient Universal Morphology Control (ICML 2024 · similarity 0.90)
- Scaling Proprioceptive-Visual Learning with Heterogeneous Pre-trained Transformers (NeurIPS 2024 · similarity 0.88)
- MeMo: Meaningful, Modular Controllers via Noise Injection (NeurIPS 2024 · similarity 0.87)
- MorphVAE: Advancing Morphological Design of Voxel-Based Soft Robots with Variational Autoencoders (AAAI 2024 · similarity 0.87)
- DittoGym: Learning to Control Soft Shape-Shifting Robots (ICLR 2024 · similarity 0.87)
- Meta-Controller: Few-Shot Imitation of Unseen Embodiments and Tasks in Continuous Control (NeurIPS 2024 · similarity 0.86)
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.