Frequently Asked Questions

General Questions

Which training method should I use?

• Regression (regression): Simplest baseline, give strong performance on most tasks with correct architecture.
• MIP (mip): Minimum Iterative Policy, reserve most of the benefits of flow matching while inference faster.
• Flow matching (flow): Good default choice, best performance on most tasks.
• CTM/PSD/LSD (ctm/psd/lsd): Self distillation, slower to train. Useful for learning full aciton distribution.

See the paper for detailed comparisons and recommendations.

Which network architecture should I use?

• State observations: mlp (simple, fast) or chiunet (better for long horizons)
• Image observations: chiunet (proven architecture) or chitransformer (for complex scenes)
• Long horizons (>16): chitransformer or sudeepdit
• Limited compute: mlp

Installation & Setup

Do I need to use UV package manager?

While UV is recommended for its speed, you can use pip instead:

pip install -e .

Can I pause and resume training?

Yes! MIP supports auto-resume:

# Training will automatically resume from last checkpoint if it exists
uv run examples/train_robomimic.py task=lift_ph_state optimization.auto_resume=true

Performance

How can I speed up training?

1. Enable torch.compile to speed up training:

   uv run examples/train_robomimic.py optimization.use_compile=true

2. Use CUDA graphs to reduce cpu overhead (state observations only):

   uv run examples/train_robomimic.py optimization.use_cudagraphs=true

You can use both to get the best performance.