so_arm100 tasks' simulation is 85% ready for PR 🙂

[Greenery-S]

✅ Description:
After 3 weeks of learning and trying, I’ve finally built the so_arm100 tasks, including So100PickCube and So100PickCubeOrientation.

🔧 What I contributed:

• Modified the original so_arm100 MJCF to simplify collision handling
• Created the task src code
• Designed the reward functions — the hardest part due to their unpredictability

🎥 Result video link:

download_better4.mp4

📅 Next step:

• Verify these tasks on the real robot in August
• Improve the reward function for better training performance
• Build two so101 arms — we might add more tasks after that!

I'm a PhD student at NUS working on offline reinforcement learning, primarily using the JAX/Flax stack for its efficiency and flexibility.
Previously, my work focused on benchmark performance (e.g., D4RL), but this is my first time designing and training a task from scratch.
It gave me a clearer understanding of the challenges in real-world task design — a valuable shift from benchmark-driven research.

Tips for beginners building custom tasks:

• Minimize simulation cost — it’s always the root cause of slow JIT compilation and NaN issues!
• Use simple geometric shapes to approximate collisions
• Prefer implicitfast integrator with minimal solver iterations
• Stick to JAX 0.4.35 or 0.5.x for better compatibility and performance

[btaba]

That's awesome @Greenery-S thank you for the post! Moving this to "discussions" for more viz.