[Fix] add_new_robot tutorial wheel-velocity action on GPU device

[hujc7]

1. Summary

• scripts/tutorials/01_assets/add_new_robot.py errored at runtime: the wheel-velocity action was a CPU torch.Tensor and set_joint_velocity_target dispatches to a GPU Warp kernel that rejects CPU tensors at type check.
• Allocate the two wheel-velocity templates once on sim.device, reuse across the loop, and call set_joint_velocity_target_index (also silences the deprecation warning visible in the bug report's first line).
• Scope: tutorial only. 1 file, +12 −3 plus a changelog fragment.

2. Why it broke

• The tutorial was added in Using assets from Isaac sim (USD) #289 (2025-03-12) with torch.Tensor([[10.0, 10.0]]) — CPU literal. The asset API at the time accepted CPU torch tensors.
• Migrating assets from torch to warp #4551 (2026-02-13) migrated the assets layer from torch to Warp kernels; CPU torch tensors silently became unsupported because Warp's kernel-argument check requires GPU-accessible data.
• Fixes missing warp to torch conversions in demos and tutorials #4911 (2026-03-10) followed up with a sweep that fixed several tutorials. It touched the reset block of this file (added wp.to_torch and _index variants for write_*_to_sim), but did not touch the per-step action block — that remained CPU.

The bug has therefore been latent since #4551 and surfaced now that more users are running the tutorial on v3.0.0-beta.

3. Why only this tutorial

A grep across scripts/tutorials/**.py for tensor literals fed into set_joint_*_target* / write_*_to_sim_index shows add_new_robot.py is the only file constructing tensors without a device= kwarg. Every other tutorial either reads from a .torch accessor (already on device) or uses torch.tensor(..., device=sim.device).

4. Test plan

• Reproduced the customer's error in isolation: CPU torch.Tensor → wp.launch → argument 'in_data' expects an array of type array(ndim=2, dtype=float32), but passed value has type Tensor.
• Same launch with a GPU tensor (.cuda()) succeeds.
• Pre-commit clean on changed files.
• Manual smoke: ./isaaclab.sh -p scripts/tutorials/01_assets/add_new_robot.py --num_envs 1 --headless runs through several reset cycles.

5. Out of scope

• The underlying API in isaaclab_physx silently accepts CPU torch tensors at the Python boundary and only fails inside wp.launch. Auto-moving CPU tensors to self.device in set_joint_*_target_index would prevent users from hitting this from other scripts. Not in scope here per the bug-report preference for a tutorial-only fix; flagging for follow-up.

Refs NVBug 6156303

[isaaclab-review-bot]

Code Review Summary

This PR correctly fixes a runtime error in the add_new_robot.py tutorial caused by passing CPU tensors to GPU Warp kernels. The fix is well-structured and follows established patterns in other Isaac Lab tutorials.

Strengths

• Root cause addressed: Pre-allocating tensors on sim.device ensures compatibility with the GPU-based asset layer.
• API modernization: Migrating from the deprecated set_joint_velocity_target() to set_joint_velocity_target_index() addresses the deprecation warning noted in the bug report.
• Efficient pattern: Allocating tensors once outside the loop rather than per-iteration is the correct approach.
• Thorough PR description: The root cause analysis and historical context (PRs #289, #4551, #4911) help reviewers understand why this bug existed and why only this tutorial was affected.

Minor Observations

No blocking issues found. The implementation is clean and matches the patterns used in other tutorials (e.g., run_articulation.py).

CI Note

The "Check for Broken Links" CI failure is unrelated to this PR — it's caused by an external site (threedworld.org) returning HTTP 403.

LGTM from a code perspective.

[isaaclab-review-bot]

Good practice: pre-allocating these tensors on the correct device outside the loop avoids per-iteration allocation overhead and ensures GPU compatibility.