Newton: push_by_setting_velocity causes training divergence (angular velocity frame mismatch)

[ax-anoop]

Bug

push_by_setting_velocity causes 100% fall rate and training divergence when using the Newton physics backend (presets=newton). The same push magnitudes work correctly with PhysX.

Root Cause

push_by_setting_velocity calls asset.write_root_velocity_to_sim_index() which delegates to write_root_com_velocity_to_sim_index() via a warp kernel. This kernel appears to not correctly convert world-frame angular velocity to MuJoCo's body-frame qvel convention.

MuJoCo stores free joint velocities as [vx_world, vy_world, vz_world, wx_body, wy_body, wz_body] — the angular part is in body frame. The warp kernel appears to write world-frame angular velocity directly, creating physically incorrect torques that destabilize the robot.

Evidence: The equivalent push in mjlab (which uses the same MuJoCo physics) works correctly because it manually converts angular velocity:

# mjlab's working implementation:
ang_vel_b = quat_apply_inverse(quat_w, velocity[:, 3:])
velocity_qvel = torch.cat([velocity[:, :3], ang_vel_b], dim=-1)
self.data.qvel[env_ids, free_joint_v_adr] = velocity_qvel

Workaround

Bypass write_root_velocity_to_sim_index and write directly to model.joint_qd with manual frame conversion:

root_quat_w = wp.to_torch(asset.data.root_link_quat_w)[env_ids]
ang_vel_b = math_utils.quat_apply(math_utils.quat_conjugate(root_quat_w), vel_w[:, 3:])
joint_qd = wp.to_torch(model.joint_qd).reshape(num_envs, -1)
joint_qd[env_ids, :3] = vel_w[:, :3]   # linear (world frame)
joint_qd[env_ids, 3:6] = ang_vel_b     # angular (body frame)

Results

Push method	1K iter reward (4096 envs)
Stock push_by_setting_velocity	-1028 (diverged, 100% falls)
Direct qvel write with frame conversion	+16.07 (converges)
PhysX stock push (reference)	+22.36

Reproduction

python scripts/train.py \
  --task Isaac-Velocity-Flat-H1-v0 \
  --num_envs 4096 --max_iterations 200 \
  presets=newton

The robot will start standing, then fall over by iteration 10 and never recover.

Environment

• IsaacLab v3.0.0-beta
• Isaac Sim 6.0 (from source, v6.0.0-dev2)
• Newton/MuJoCo-Warp backend