Support custom robot URDFs in simulation environments and layout generation

[Charcot-NUDT]

Summary

First of all, thank you for this impressive project! I would like to use
EmbodiedGen to collect training data for my own robot, but the current
pipeline appears to only support robots registered in ManiSkill's
REGISTERED_AGENTS (defaulting to Franka Panda), with no documented way
to plug in a custom robot URDF.

Use case

I have my own robot (URDF file) and would like to:

1. Load my robot into the simulation environments instead of the default
   panda — e.g., in PickEmbodiedGen-v1 and the sim-cli / parallel_sim.py
   pipelines.
2. Generate custom scenes and task layouts around my robot (via layout-cli
   or a hand-written layout.json).
3. Collect demonstration / trajectory data with my robot for policy training.

Current behavior

From reading the code:

• embodied_gen/utils/enum.py — RobotItemEnum only enumerates
  franka / ur5 / piper.
• embodied_gen/utils/simulation.py — load_mani_skill_robot instantiates
  robots via ManiSkill's REGISTERED_AGENTS; if the robot name in the layout
  is not registered, it silently falls back to panda, which is easy to miss.
• embodied_gen/envs/pick_embodiedgen.py — PickEmbodiedGen-v1 is the only
  task environment, hard-coded with robot_uids="panda".
• The documentation (tutorials) does not cover how to integrate a custom robot.

Feature requests

1. A documented, first-class way to register a custom robot URDF
   (e.g., a --robot_urdf argument, a robot config file, or a tutorial on
   writing a minimal ManiSkill BaseAgent wrapper), usable across
   sim-cli, parallel_sim.py, and the gym environments.
2. An explicit error or warning instead of silently falling back to
   panda when the requested robot is not found.
3. Optionally: guidance on how to define custom task environments
   (beyond pick-and-place) that work with the layout/data-collection pipeline.

Alternatives considered

I understand I could register my robot as a ManiSkill agent myself and extend
BaseEnv, but an official mechanism + documentation would make EmbodiedGen
much more useful as a data-generation platform for arbitrary robots.

Environment

• EmbodiedGen version: v2.0.0
• OS: Linux

[HochCC]

Hi @Charcot-NUDT,

Thank you for the detailed feedback. We agree that custom robot support would be valuable, but it is a relatively large feature. Besides loading the URDF, it also requires controller, mesh assets and task-specific interfaces, so we currently do not have plans to provide a generic integration.

For now, we recommend registering your robot as a custom ManiSkill BaseAgent, including its URDF and controllers, and then using it with EmbodiedGen-generated scenes through gym.make.

You are also right that silently falling back to Panda is easy to miss. We will consider replacing it with an explicit warning or error.