Publish Parkour qwenoft latency-aware PPO

latency-aware/extreme-parkour-go2/small-policy/ppo-qwenoft-rtx3090-v1/README.md

@@ -0,0 +1,20 @@
+# Extreme Parkour Go2 — QwenOFT latency-aware PPO
+
+Native command policy trained from scratch with the QwenOFT RTX 3090 temporal
+latency profile. The checkpoint completed 3,000 PPO iterations on one H200.
+
+Training used 6,144 environments, seed 1, 50 Hz control, 10 Hz decisions,
+24 decisions/environment/iteration, control repeat 5, five PPO epochs, four
+minibatches, and fixed learning rate `2e-4`.
+
+| Evaluation | Episodes | Return mean ± SD | Raw length mean ± SD | Waypoint progress mean ± SD |
+|---|---:|---:|---:|---:|
+| QwenOFT temporal profile | 100 | 16.2985 ± 5.8936 | 650.84 ± 161.56 | 0.8314 ± 0.2773 |
+| 0 ms latency | 100 | 16.7771 ± 6.1831 | 663.42 ± 173.64 | 0.8329 ± 0.2872 |
+
+Both evaluations use 256 parallel environments, seed 1, deterministic policy
+means, and a 1,500 raw-control-frame episode cap. See `eval_profile.json` and
+`eval_l0.json` for complete metrics and latency counts.
+
+Training run: https://wandb.ai/saberrr-zju/isaaclab-parkour/runs/6ix3w954
+

latency-aware/extreme-parkour-go2/small-policy/ppo-qwenoft-rtx3090-v1/checkpoint.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1c56e4cd2d6b6618c0592192dd5019e3acdc444965fc29980abbcc31efaca013
+size 239286587

latency-aware/extreme-parkour-go2/small-policy/ppo-qwenoft-rtx3090-v1/eval_l0.json

@@ -0,0 +1,44 @@
+{
+  "identity": {
+    "checkpoint": "/home/ubuntu/latency-sensitive-bench/third_party/Isaaclab_Parkour/logs/rsl_rl/unitree_go2_parkour/2026-09-16_09-14-43_extreme_parkour_go2_qwenoft_scratch_seed1_3k_h200/model_2999.pt",
+    "latency_config": "/home/ubuntu/latency-sensitive-bench/configs/training/latency/parkour_zero.yaml",
+    "latency": {
+      "method": "fixed",
+      "env_fps": 50,
+      "obs_fps": 10,
+      "profile_path": null
+    },
+    "teacher_checkpoint": "/home/ubuntu/model_14999.pt",
+    "teacher_baseline": false,
+    "seed": 1
+  },
+  "evaluation": {
+    "num_envs": 256,
+    "completed_episodes": 100,
+    "requested_episodes": 100,
+    "max_episode_steps": 1500,
+    "control_repeat": 5
+  },
+  "metrics": {
+    "reward_mean": 16.77713394165039,
+    "reward_std": 6.183106310446976,
+    "episode_length_raw_mean": 663.42,
+    "episode_length_raw_std": 173.63512073170898,
+    "waypoints_mean": 0.8328571428571427,
+    "waypoints_std": 0.28717839154683733,
+    "edge_violation_mean": 0.158,
+    "edge_violation_std": 0.415975960843893
+  },
+  "latency": {
+    "scope": "all_vector_slots_until_requested_episode_count",
+    "admissions": 46080,
+    "drops": 0,
+    "raw_control_frames": 230197,
+    "sampled_ms_mean": 0.0
+  },
+  "throughput": {
+    "raw_control_frames": 230197,
+    "elapsed_seconds": 47.57120870301151,
+    "raw_control_frames_per_second": 4838.998341142156
+  }
+}

latency-aware/extreme-parkour-go2/small-policy/ppo-qwenoft-rtx3090-v1/eval_l0.yaml

@@ -0,0 +1,12 @@
+__base__: train.yaml
+latency:
+  method: fixed
+  fixed_latency_ms: 0
+  seed: 0
+evaluation:
+  num_envs: 256
+  episodes: 100
+  max_episode_steps: 1500
+  seed: 1
+  deterministic: true
+

latency-aware/extreme-parkour-go2/small-policy/ppo-qwenoft-rtx3090-v1/eval_profile.json

@@ -0,0 +1,44 @@
+{
+  "identity": {
+    "checkpoint": "/home/ubuntu/latency-sensitive-bench/third_party/Isaaclab_Parkour/logs/rsl_rl/unitree_go2_parkour/2026-09-16_09-14-43_extreme_parkour_go2_qwenoft_scratch_seed1_3k_h200/model_2999.pt",
+    "latency_config": "/home/ubuntu/latency-sensitive-bench/outputs/latency_rl/parkour/h200_20260916_qwenoft/latency.yaml",
+    "latency": {
+      "method": "temporal",
+      "env_fps": 50,
+      "obs_fps": 10,
+      "profile_path": "/home/ubuntu/latency-sensitive-bench/results/profiling/profiles/qwenoft/1x-rtx3090/extreme_parkour_go2/instance_6e332477bffea7da/profile.json"
+    },
+    "teacher_checkpoint": "/home/ubuntu/model_14999.pt",
+    "teacher_baseline": false,
+    "seed": 1
+  },
+  "evaluation": {
+    "num_envs": 256,
+    "completed_episodes": 100,
+    "requested_episodes": 100,
+    "max_episode_steps": 1500,
+    "control_repeat": 5
+  },
+  "metrics": {
+    "reward_mean": 16.29850423336029,
+    "reward_std": 5.893599424632169,
+    "episode_length_raw_mean": 650.84,
+    "episode_length_raw_std": 161.5623043343285,
+    "waypoints_mean": 0.8314285714285714,
+    "waypoints_std": 0.2772900934371211,
+    "edge_violation_mean": 0.122,
+    "edge_violation_std": 0.3392874887171644
+  },
+  "latency": {
+    "scope": "all_vector_slots_until_requested_episode_count",
+    "admissions": 42425,
+    "drops": 3143,
+    "raw_control_frames": 227629,
+    "sampled_ms_mean": 101.2792230542913
+  },
+  "throughput": {
+    "raw_control_frames": 227629,
+    "elapsed_seconds": 47.871054112998536,
+    "raw_control_frames_per_second": 4755.0446552249905
+  }
+}

latency-aware/extreme-parkour-go2/small-policy/ppo-qwenoft-rtx3090-v1/eval_profile.yaml

@@ -0,0 +1,8 @@
+__base__: train.yaml
+evaluation:
+  num_envs: 256
+  episodes: 100
+  max_episode_steps: 1500
+  seed: 1
+  deterministic: true
+

latency-aware/extreme-parkour-go2/small-policy/ppo-qwenoft-rtx3090-v1/provenance.json

@@ -0,0 +1,34 @@
+{
+  "task": "extreme-parkour-go2",
+  "model": "ppo-command-policy",
+  "latency_profile_model": "qwenoft",
+  "status": "evaluated",
+  "wandb_url": "https://wandb.ai/saberrr-zju/isaaclab-parkour/runs/6ix3w954",
+  "checkpoint": "checkpoint.pt",
+  "checkpoint_bytes": 239286587,
+  "checkpoint_sha256": "1c56e4cd2d6b6618c0592192dd5019e3acdc444965fc29980abbcc31efaca013",
+  "checkpoint_iteration": 2999,
+  "decision_timesteps": 442368000,
+  "control_frames": 2211840000,
+  "initialization": "from_scratch",
+  "training_seed": 1,
+  "source_revisions": {
+    "benchmark": "f0929454423a5f170eb3c226aec42103ea351248",
+    "isaac_lab": "46dff135f44683f031edf346e544fcfd8456b2bb",
+    "parkour": "b34830ece802978aaff8b97500a3a924d1209c21"
+  },
+  "teacher": {
+    "checkpoint_sha256": "acd97fbc316cdfa52521f05ee6c6be8614a18b28cb9ff918a4aa9e25c7e29378"
+  },
+  "profile": {
+    "repo_id": "latency-sensitive-bench/profiles",
+    "revision": "e8be1e5dfd309f85ab7ff704dc02bd4898b7ceb4",
+    "path": "profiles/qwenoft/1x-rtx3090/extreme_parkour_go2/instance_6e332477bffea7da",
+    "profile_sha256": "5059d7f6acf2289299d8f25b04e295cad9823be317f49186e94599a4dcf337e0"
+  },
+  "evaluations": {
+    "profile": "eval_profile.json",
+    "zero_latency": "eval_l0.json"
+  }
+}
+

latency-aware/extreme-parkour-go2/small-policy/ppo-qwenoft-rtx3090-v1/train.yaml

@@ -0,0 +1,34 @@
+experiment:
+  name: extreme_parkour_go2_qwenoft_scratch
+  seed: 1
+env:
+  name: extreme_parkour_go2
+  env_fps: 50
+  obs_fps: 10
+  simulator_device: cuda:0
+latency:
+  method: temporal
+  seed: 0
+  profile_path: profiles/qwenoft/1x-rtx3090/extreme_parkour_go2/instance_6e332477bffea7da/profile.json
+executor:
+  simulated_worker_capacity: 1
+scheduler:
+  hold_policy: hold
+  ordering_policy: latest_ready
+command:
+  dimension: 34
+  horizon: 1
+  control_repeat: 5
+training:
+  num_envs: 6144
+  kit_threads: 8
+  physx_threads: 8
+  physics_dt: 0.005
+  decimation: 4
+  decisions_per_env_per_iteration: 24
+  max_iterations: 3000
+  learning_epochs: 5
+  minibatches: 4
+  learning_rate: 0.0002
+  learning_rate_schedule: fixed
+