diff --git a/reproduction/asap-isaacgym-conda.yaml b/reproduction/asap-isaacgym-conda.yaml
new file mode 100644
index 0000000..c94d018
--- /dev/null
+++ b/reproduction/asap-isaacgym-conda.yaml
@@ -0,0 +1,6 @@
+name: asap
+channels:
+  - defaults
+dependencies:
+  - python=3.8
+  - pip
diff --git a/reproduction/isaacgym-smoke-report.md b/reproduction/isaacgym-smoke-report.md
new file mode 100644
index 0000000..81fb761
--- /dev/null
+++ b/reproduction/isaacgym-smoke-report.md
@@ -0,0 +1,111 @@
+# ASAP Isaac Gym smoke reproduction
+
+## Goal
+
+Verify that the upstream G1 locomotion stack can initialize on GPU without
+running a baseline training iteration. This is a pre-migration baseline for a
+future R1 robot port.
+
+## Environment
+
+- Repository revision: `df5320c`
+- Conda environment: `asap`
+- Python: 3.8.20
+- PyTorch: 2.4.1+cu121
+- Torchvision: 0.19.1+cu121
+- Isaac Gym: Preview 4 (`1.0rc4` package metadata)
+- GPU used for verification: NVIDIA GeForce RTX 4090
+- NVIDIA driver: 580.76.05
+
+Create the environment:
+
+```bash
+conda env create -f reproduction/asap-isaacgym-conda.yaml
+conda activate asap
+
+pip install torch==2.4.1 torchvision==0.19.1 \
+  --index-url https://download.pytorch.org/whl/cu121
+pip install -e /path/to/isaacgym/python --no-deps
+pip install scipy==1.10.1 pyyaml imageio ninja
+pip install -e . -e isaac_utils
+```
+
+Isaac Gym Preview 4 can be downloaded separately with:
+
+```bash
+curl -L -o IsaacGym_Preview_4_Package.tar.gz \
+  https://developer.nvidia.com/isaac-gym-preview-4
+```
+
+Keep the archive and extracted SDK outside this repository.
+
+## Verification
+
+Dependency consistency:
+
+```bash
+python -m pip check
+```
+
+Result: `No broken requirements found.`
+
+The Isaac Gym GPU PhysX smoke test created a sphere on a plane, enabled the GPU
+pipeline, advanced 120 simulation steps, and read the final state through the
+tensor API. Result:
+
+```text
+ISAAC_GYM_SMOKE_OK NVIDIA GeForce RTX 4090 steps=120 z=0.102
+```
+
+The ASAP G1 stack was initialized without training using:
+
+```bash
+CUDA_DEVICE_ORDER=PCI_BUS_ID CUDA_VISIBLE_DEVICES=4 \
+HYDRA_FULL_ERROR=1 python humanoidverse/train_agent.py \
+  +simulator=isaacgym \
+  +exp=locomotion \
+  +domain_rand=NO_domain_rand \
+  +rewards=loco/reward_g1_locomotion \
+  +robot=g1/g1_29dof_anneal_23dof \
+  +terrain=terrain_locomotion_plane \
+  +obs=loco/leggedloco_obs_singlestep_withlinvel \
+  num_envs=1 \
+  project_name=ASAPSmoke \
+  experiment_name=G1InitOnly \
+  headless=True \
+  use_wandb=False \
+  algo.config.num_learning_iterations=0
+```
+
+Result: one G1 environment was created with GPU PhysX and the GPU pipeline.
+ASAP resolved an 81-dimensional actor observation, an 81-dimensional critic
+observation, and 23 actions. PPO and rollout storage initialized successfully.
+No learning iteration ran.
+
+The bundled ONNX models also load with ONNX Runtime:
+
+- Decoupled locomotion: input `[1, 500]`, output `[1, 12]`
+- CR7 mimic: input `[1, 380]`, output `[1, 23]`
+
+## Local outputs
+
+- Isaac Gym SDK: `/home/HDD/czc/deps/asap/isaacgym`
+- Isaac Gym archive: `/home/HDD/czc/deps/asap/IsaacGym_Preview_4_Package.tar.gz`
+- Initialization output: `logs/ASAPSmoke/20260716_072756-G1InitOnly-locomotion-g1_29dof_anneal_23dof`
+
+The output directory is ignored by Git. It contains the resolved config, short
+logs, TensorBoard metadata, and an untrained 1.7 MB `model_0.pt` emitted by the
+upstream zero-iteration code path.
+
+## Limits and R1 migration notes
+
+- Isaac Gym Preview 4 and this Python 3.8 stack do not support RTX 5090
+  (`sm_120`). Use an RTX 4090 (`sm_89`) for this baseline.
+- This stage verifies simulator and G1 environment initialization, not policy
+  quality or baseline training.
+- The current robot contract is coupled to the G1 URDF/MJCF assets, 29-DoF
+  state, 23 controlled joints, joint/body names, default pose, PD gains, limits,
+  observation dimensions, reward contact bodies, and ONNX action dimensions.
+- An R1 port should first add an R1 robot/asset config and a deterministic joint
+  mapping, then validate reset and zero-action stepping before adapting rewards
+  or reusing policy checkpoints.