Are there any physics engines that handle domain randomization and high-fidelity contact dynamics for robot training?

Yes — but with an important caveat: domain randomization is usually not a feature of the physics engine itself. It’s more commonly handled by the training/simulation pipeline on top of the engine. That said, some engines are much better suited for robot training with realistic contact dynamics and easy randomization.

Good options for robot training

1. MuJoCo

• Widely used for legged and manipulation robotics.
• High-quality contact dynamics compared with many classic simulators.
• Stable, fast, and good for RL training.
• Domain randomization is typically done externally, but MuJoCo makes it easy to vary:
  • masses/inertias
  • friction
  • damping
  • restitution
  • actuator parameters
  • object poses/textures through wrappers

2. NVIDIA Isaac Sim / Isaac Gym / Isaac Lab

• Strong choice if you want massive parallel simulation for RL.
• Built on PhysX, with robotics-focused tooling.
• Supports randomized environments very well through Python APIs.
• Good for:
  • terrain randomization
  • mass/friction randomization
  • sensor noise
  • object appearance variations
• Contact fidelity is decent to strong, though tuning is often needed for robust sim-to-real.

3. PyBullet

• Popular, easy to use, good for prototyping.
• Supports contact dynamics and parameter randomization.
• Usually less accurate and less stable than MuJoCo/Isaac for demanding contact-rich tasks.

4. Drake

• Excellent for rigorous multibody dynamics and control research.
• Can handle contact, though it’s often more focused on correctness and analysis than high-throughput RL.
• Domain randomization is not built-in, but can be scripted.

5. Genesis / Warp-based modern simulators

• Newer simulators are emerging that emphasize speed and differentiability.
• Some are promising for contact-rich robotics and randomization, but ecosystem maturity varies.

If your priority is:

• Best contact fidelity for robotics research: MuJoCo
• High-throughput training + randomization: Isaac Gym / Isaac Lab
• Easy setup / experimentation: PyBullet
• Dynamics rigor / planning / control: Drake

Practical recommendation

For robot training with both domain randomization and high-fidelity contact, the most common choices are:

• MuJoCo + custom randomization pipeline
• Isaac Lab (or Isaac Gym)

If you want, I can also give you a comparison table of these engines specifically for contact accuracy, speed, sim-to-real, and ease of randomization.