Research map / Reinforcement learning / Visual and representation RL
OXE-AugE: A Large-Scale Robot Augmentation of OXE for Scaling Cross-Embodiment Policy Learning
Guanhua Ji, Harsha Polavaram, Lawrence Yunliang Chen, Sandeep Bajamahal, Zehan Ma, Simeon Adebola, Chenfeng Xu, Ken Goldberg
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| add augmented robots (N) to training data (ablation) | Diffusion Policy on Franka source robot · Robosuite 5 tasks, visual perturbations (lighting, occlusion) | success rate | Improved |
| add source data to single-robot augmented data (ablation) | Diffusion Policy, augmented target robots · Robosuite 5 tasks, transfer | success rate: +9-19% | Improved |
| scale from 1 to N augmented robots for already-seen target robots (ablation) | Diffusion Policy, augmented target robots · Robosuite 5 tasks, transfer | success rate | Improved |
| augment with N-1 other robots instead of none (ablation) | Diffusion Policy, unseen robots · Robosuite 5 tasks, generalization | success rate | Improved |
| more augmented robots with only 25% of source trajectories vs fewer robots with 100% (ablation) | Diffusion Policy, unseen robots · Robosuite 5 tasks, source data scaling | success rate | Improved |
| use simulation-based AugE-Toolkit instead of Shadow | Diffusion Policy · Robosuite transfer | success rate: +25% on average | Improved |
| use AugE-Toolkit instead of Shadow | Diffusion Policy · Robosuite unseen-robot generalization | success rate | No clear effect |
| use wrist-view only policies instead of third-person with augmentation | Diffusion Policy · Robosuite transfer | success rate | Hurt |
| fine-tune on OXE-AugE with N augmented robots instead of Bridge only | OpenVLA-OFT · Real-world Bridge tasks, Franka and Robotiq++ | success rate: +24% | Improved |
| fine-tune on OXE-AugE with N augmented robots instead of Bridge only | pi0 · Real-world Bridge tasks, Franka and Robotiq++ | success rate: +45% | Improved |
| fine-tune on 2 augmented robots (Franka, UR5e) instead of none (ablation) | OpenVLA-OFT and pi0 · Real-world Bridge tasks | success rate | Improved |
| use simulation-rendered augmentation instead of diffusion-based RoVi-Aug (ablation) | fine-tuned policies on Bridge · Real-world Bridge tasks | success rate: +48-77% | Improved |
Closest papers by meaning
- Scaling Proprioceptive-Visual Learning with Heterogeneous Pre-trained Transformers (NeurIPS 2024 · similarity 0.89)
- Data Scaling Laws in Imitation Learning for Robotic Manipulation (ICLR 2025 · similarity 0.88)
- DexScale: Automating Data Scaling for Sim2Real Generalizable Robot Control (ICML 2025 · similarity 0.88)
- STRAP: Robot Sub-Trajectory Retrieval for Augmented Policy Learning (ICLR 2025 · similarity 0.88)
- Meta-Controller: Few-Shot Imitation of Unseen Embodiments and Tasks in Continuous Control (NeurIPS 2024 · similarity 0.88)
- What's the Move? Hybrid Imitation Learning via Salient Points (ICLR 2025 · similarity 0.87)
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.
