SWAP: Symmetric Equivariant World-Model for Agile Robot Parkour
TLDR
SWAP embeds symmetry into a world model for quadruped parkour, achieving record-breaking jumps and zero-shot generalization.
Reasoning
The paper presents a novel symmetric equivariant world model that reduces learning burden and achieves impressive real-world parkour records. Strengths include real-world validation and strong geometric generalization, but the approach is limited to symmetric quadruped locomotion and lacks explicit baseline comparisons in the abstract.
Read-first score
Read-first score 53.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 35.
Field roles
Rank sensitivity
Stability: volatile; rank range: 343.
Keyword Scores
Deep Analysis
Innovations
- End-to-end equivariant symmetric world model that embeds left-right symmetry directly into both the world model and actor-critic networks
- Reduction of learning burden by avoiding redundant encoding of symmetric interactions in latent world models
- Demonstration that symmetry equivariance serves as an effective structural prior for pushing the physical boundaries of learned legged locomotion
Methodology
SWAP is an end-to-end equivariant symmetric world model that incorporates symmetry directly into the world model and actor-critic networks. It uses a latent world model architecture with equivariant layers to enforce left-right symmetry, reducing redundant learning. The model is trained on real-world robot data and evaluated on parkour tasks including gap crossing and platform climbing.
Key Results
The robot successfully leaps across a 2.13 m gap and climbs a 1.63 m platform, breaking previous records for quadruped parkour. The framework also demonstrates robust geometric generalization to unseen mirrored terrains and zero-shot transferability across diverse outdoor environments.