docs: record Isaac Gym smoke reproduction

[Zarato2122]

Stage goal

Document a verified, zero-training-iteration Isaac Gym setup for the upstream G1 locomotion stack as a baseline for a future R1 port.

Changed files

• reproduction/asap-isaacgym-conda.yaml
• reproduction/isaacgym-smoke-report.md

Test commands

• python -m pip check
• GPU PhysX sphere smoke test with 120 simulation steps
• humanoidverse/train_agent.py with num_envs=1 and algo.config.num_learning_iterations=0
• ONNX Runtime load checks for the bundled locomotion and CR7 models

Test results

• Dependencies consistent
• GPU PhysX and GPU tensor pipeline passed on RTX 4090
• One G1 environment initialized successfully
• Actor/critic observations: 81/81; actions: 23
• PPO and rollout storage initialized; no learning iteration ran
• Bundled ONNX models loaded successfully

Local output paths

• logs/ASAPSmoke/20260716_072756-G1InitOnly-locomotion-g1_29dof_anneal_23dof
• /home/HDD/czc/deps/asap/isaacgym

Uncommitted large files

• /home/HDD/czc/deps/asap/IsaacGym_Preview_4_Package.tar.gz
• Generated logs and untrained model_0.pt under the ignored local output path

Current limits and next step

Isaac Gym Preview 4 does not support RTX 5090 sm_120 in this stack, so validation used RTX 4090. This stage does not assess policy quality or train a baseline. The R1 port should next add R1 assets/configuration and deterministic joint/body mappings, then validate reset and zero-action stepping before adapting rewards or checkpoints.