Research map / Reinforcement learning / Visual and representation RL
RoboCasa365: A Large-Scale Simulation Framework for Training and Benchmarking Generalist Robots
Soroush Nasiriany, Sepehr Nasiriany, Abhiram Maddukuri, Yuke Zhu
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 |
|---|---|---|---|
| use GR00T N1.5 instead of Diffusion Policy | multi-task policy · 300-task multi-task average | success rate: 6.1 -> 20.0 | Improved |
| pretrain on RoboCasa365 data before target fine-tuning (100% target) (ablation) | GR00T N1.5 · target average | success rate: 43.7 -> 51.1 | Improved |
| pretrain before fine-tuning with 10% target data (ablation) | GR00T N1.5 · target average | success rate: 21.0 -> 35.9 | Improved |
| add MimicGen synthetic data to human pretraining (10% target) (ablation) | GR00T N1.5 · target average | success rate: 40.0 -> 35.9 | Hurt |
| add MimicGen synthetic data to human pretraining (100% target) (ablation) | GR00T N1.5 · target average | success rate: 52.5 -> 51.1 | Hurt |
| increase pretraining tasks from 50 to 300 (10% target) (ablation) | GR00T N1.5 · target average | success rate: 34.7 -> 40.0 | Improved |
| increase pretraining scenes from 5 to 2500 (ablation) | GR00T N1.5 · zero-shot target kitchens | success rate: 29.6 -> 44.7 | Improved |
| use full fine-tuning instead of LoRA (ablation) | GR00T N1.5 · multi-task average | success rate: 1.2 -> 20.0 | Improved |
| use two-stage pretrain then fine-tune instead of joint co-training (ablation) | GR00T N1.5 · target average | success rate: 22.5 -> 51.1 | Improved |
| add sim data to real data for real-robot training | GR00T N1.5 · real-world 4 tasks | success rate: 61.8 -> 79.8 | Improved |
| learn longer-horizon tasks sequentially | GR00T N1.5 · lifelong learning atomic tasks | success rate: 41.5 -> 10.6 | Hurt |
| perturb camera pose at test time (ablation) | GR00T N1.5 · Composite-Seen | success rate: 40.6 -> 28.8 | Hurt |
Closest papers by meaning
- GenSim: Generating Robotic Simulation Tasks via Large Language Models (ICLR 2024 · similarity 0.89)
- RoboCerebra: A Large-scale Benchmark for Long-horizon Robotic Manipulation Evaluation (NeurIPS 2025 · similarity 0.89)
- Zero-Shot Robotic Manipulation with Pre-Trained Image-Editing Diffusion Models (ICLR 2024 · similarity 0.88)
- GENMANIP: LLM-driven Simulation for Generalizable Instruction-Following Manipulation (CVPR 2025 · similarity 0.88)
- RoboAlign-R1: Distilled Multimodal Reward Alignment for Robot Video World Models (NeurIPS 2026 · similarity 0.88)
- RoboGen: Towards Unleashing Infinite Data for Automated Robot Learning via Generative Simulation (ICML 2024 · similarity 0.88)
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.
