Publish MIKASA PPO trained with QwenOFT RTX 3090 latency profile

README.md

@@ -3,10 +3,11 @@ license: other
 ---
 # Latency-Sensitive Bench models
 
-Inference-ready zero-latency teachers and VLA policies for the supported benchmark tasks.
+Inference-ready teachers and VLA policies for the supported benchmark tasks.
 
 ## Layout
 
+- `latency-aware/mikasa-intercept-grab-fast/small-policy/ppo-qwenoft-rtx3090-v1`: scratch-trained PPO with the QwenOFT RTX 3090 temporal profile; independent 256-episode final success 98.83% (253/256).
 - `zero-latency/mikasa-intercept-grab-fast/small-policy/ppo-v1`: accepted PPO teacher and strict evaluation.
 - `zero-latency/mikasa-intercept-grab-fast/vla/starvla-qwenoft-v1`: StarVLA QwenOFT 5k inference checkpoint, portable config, statistics, and zero-latency evaluation.
 - `zero-latency/humanoidbench-balance-simple/small-policy/fasttd3-v1`: accepted FastTD3 teacher and portable train/eval configs.

latency-aware/mikasa-intercept-grab-fast/small-policy/ppo-qwenoft-rtx3090-v1/README.md

@@ -0,0 +1,87 @@
+# MIKASA InterceptGrabFast — PPO with QwenOFT RTX 3090 latency
+
+A small privileged-state PPO policy trained from random initialization with the
+temporal latency profile measured for QwenOFT on one RTX 3090. The policy uses
+49D state observations and deterministic mean actions for evaluation.
+
+## Files
+
+| File | Contents |
+| --- | --- |
+| `checkpoint.pt` | Inference-ready native MIKASA `AgentStateOnly` state dictionary. |
+| `strict-eval.json` | Independent 256-episode evaluation under the temporal profile. |
+| `train.yaml` | Portable scratch-training configuration. |
+| `eval.yaml` | Portable configuration reproducing the independent evaluation. |
+| `provenance.json` | Selected training step, W&B run, and pinned profile source. |
+
+## Training and latency
+
+- Environment: `InterceptGrabFast-VLA-v0`, `normalized_dense` reward, 60-step episodes.
+- Observation: 49D privileged state; actor and critic each use three 256-unit tanh layers.
+- Action: 7D `pd_ee_delta_pose`, one action per accepted request.
+- Control frequency: 20 Hz; physics frequency: 100 Hz.
+- Seed: 123; 1,024 environments; 60-step rollouts; 32 minibatches; eight update epochs.
+- Learning rate: `1e-4`; PPO target KL: `0.2`; gamma: `0.99`; GAE lambda: `0.95`.
+- Latency: `temporal`, training latency seed 0, one simulated worker,
+  `latest_ready` ordering, and `hold` between action arrivals.
+- The profile's mean observation-to-action latency is 99.05 ms. The joint model
+  represents both worker occupancy and action-ready time.
+- Selected checkpoint: 13,824,000 environment steps, iteration 225, chosen by
+  periodic `eval/success_at_end`. `train.yaml` retains the original 150M training
+  budget; the selected checkpoint is from the step stated above.
+
+Training run: [ks6d837t](https://wandb.ai/saberrr-zju/MIKASA-Robo-dataset-collectors/runs/ks6d837t).
+
+The exact latency profile is
+[`profiles/qwenoft/1x-rtx3090/mikasa_intercept_grab_fast/instance_3a0d42681a03715c`](https://huggingface.co/datasets/latency-sensitive-bench/profiles/tree/66414463f2a609316da63935941a79abf0a73ce9/profiles/qwenoft/1x-rtx3090/mikasa_intercept_grab_fast/instance_3a0d42681a03715c),
+pinned to revision `66414463f2a609316da63935941a79abf0a73ce9`.
+Latency is simulated from this profile during small-policy training and evaluation.
+
+## Independent evaluation
+
+The evaluation used 64 parallel environments, 240 control steps per environment
+(four complete episodes each), environment seed `20260910`, and independent
+latency seed `271828`. Actions were deterministic policy means. The seed is a
+vector-environment reset seed; it is not an enumerated list of 256 episode seeds.
+
+| Metric | Result |
+| --- | --- |
+| Success at least once | 255 / 256 = **99.61%** |
+| Success at the end | 253 / 256 = **98.83%** |
+| Mean episode return | 33.6446 |
+| Mean episode length | 60 control steps |
+| Accepted / capacity-dropped requests | 7,606 / 7,754 |
+
+`success_at_end` is the primary metric: the ball is grasped and the robot is
+static at the final control step. This checkpoint passes the 90% acceptance
+target. Per-episode outcomes were not retained by this evaluation; the JSON
+contains the recorded aggregate results.
+
+## Download and evaluate
+
+Run from a benchmark checkout with its MIKASA environment installed:
+
+```bash
+BENCH_ROOT="$PWD"
+MODEL_PATH="latency-aware/mikasa-intercept-grab-fast/small-policy/ppo-qwenoft-rtx3090-v1"
+PROFILE_PATH="profiles/qwenoft/1x-rtx3090/mikasa_intercept_grab_fast/instance_3a0d42681a03715c"
+BUNDLE="$BENCH_ROOT/outputs/published-models/$MODEL_PATH"
+
+hf download latency-sensitive-bench/benchmark-models \
+  --include "$MODEL_PATH/*" --local-dir "$BENCH_ROOT/outputs/published-models"
+hf download latency-sensitive-bench/profiles --repo-type dataset \
+  --revision 66414463f2a609316da63935941a79abf0a73ce9 \
+  --include "$PROFILE_PATH/*" --local-dir "$BUNDLE"
+
+cd "$BUNDLE"
+CUDA_VISIBLE_DEVICES=0 \
+PYTHONPATH="$BENCH_ROOT:$BENCH_ROOT/third_party/MIKASA-Robo" \
+  "$BENCH_ROOT/third_party/MIKASA-Robo/.venv/bin/python" \
+  "$BENCH_ROOT/third_party/MIKASA-Robo/mikasa_robo_suite/vla/dataset_collectors/get_dataset_collectors_ckpt.py" \
+  --env-id InterceptGrabFast-VLA-v0 \
+  --latency-config eval.yaml --latency-output-dir evaluation
+```
+
+For scratch-training reproduction, use `train.yaml` in the same command and
+choose a separate latency output directory. The benchmark integration supplies
+the native policy architecture, simulator, and action-latency wrapper.

latency-aware/mikasa-intercept-grab-fast/small-policy/ppo-qwenoft-rtx3090-v1/checkpoint.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9bc3b3875af77b664ef246ae1dc66bf3e013f5e4a1143f89c79c8fb086dcfea1
+size 1169937

latency-aware/mikasa-intercept-grab-fast/small-policy/ppo-qwenoft-rtx3090-v1/eval.yaml

@@ -0,0 +1,42 @@
+env:
+  env_fps: 20
+  obs_fps: 20
+  noop_action:
+  - 0
+  - 0
+  - 0
+  - 0
+  - 0
+  - 0
+  - 0
+executor:
+  simulated_worker_capacity: 1
+latency:
+  method: temporal
+  seed: 271828
+  profile_path: profiles/qwenoft/1x-rtx3090/mikasa_intercept_grab_fast/instance_3a0d42681a03715c/profile.json
+scheduler:
+  ordering_policy: latest_ready
+  hold_policy: hold
+training:
+  num_envs: 1
+  num_eval_envs: 64
+  num_steps: 60
+  num_minibatches: 32
+  update_epochs: 8
+  total_timesteps: 60
+  eval_freq: 25
+  track: false
+  exp_name: small-ppo-mikasa-intercept-grab-fast-qwenoft-rtx3090-temporal-scratch-baseline-settings
+  checkpoint: checkpoint.pt
+  resume: null
+  seed: 20260910
+  num_eval_steps: 240
+  learning_rate: 0.0001
+  target_kl: 0.2
+  evaluate: true
+  save_model: false
+profile_source:
+  repo_id: latency-sensitive-bench/profiles
+  revision: 66414463f2a609316da63935941a79abf0a73ce9
+  path: profiles/qwenoft/1x-rtx3090/mikasa_intercept_grab_fast/instance_3a0d42681a03715c

latency-aware/mikasa-intercept-grab-fast/small-policy/ppo-qwenoft-rtx3090-v1/provenance.json

@@ -0,0 +1,24 @@
+{
+  "task": "mikasa-intercept-grab-fast",
+  "model": "ppo-mlp-state",
+  "status": "accepted",
+  "run_id": "ks6d837t",
+  "wandb_url": "https://wandb.ai/saberrr-zju/MIKASA-Robo-dataset-collectors/runs/ks6d837t",
+  "checkpoint": "checkpoint.pt",
+  "checkpoint_bytes": 1169937,
+  "checkpoint_global_step": 13824000,
+  "checkpoint_iteration": 225,
+  "selection_metric": "eval/success_at_end",
+  "initialization": "from_scratch",
+  "training_seed": 123,
+  "profile": {
+    "repo_id": "latency-sensitive-bench/profiles",
+    "revision": "66414463f2a609316da63935941a79abf0a73ce9",
+    "path": "profiles/qwenoft/1x-rtx3090/mikasa_intercept_grab_fast/instance_3a0d42681a03715c"
+  },
+  "trainer_source": {
+    "wandb_run_id": "ks6d837t",
+    "path": "code/third_party/MIKASA-Robo/mikasa_robo_suite/vla/dataset_collectors/get_dataset_collectors_ckpt.py"
+  },
+  "evaluation": "strict-eval.json"
+}

latency-aware/mikasa-intercept-grab-fast/small-policy/ppo-qwenoft-rtx3090-v1/strict-eval.json

@@ -0,0 +1,22 @@
+{
+  "episodes": 256,
+  "metrics": {
+    "success_once": 0.99609375,
+    "return": 33.64462661743164,
+    "episode_len": 60.0,
+    "reward": 0.5607438087463379,
+    "success_at_end": 0.98828125
+  },
+  "admitted_count": 7606,
+  "dropped_count": 7754,
+  "checkpoint": "checkpoint.pt",
+  "checkpoint_global_step": 13824000,
+  "evaluation_config": "eval.yaml",
+  "success_rate": 0.98828125,
+  "success_metric": "success_at_end",
+  "success_once_count": 255,
+  "success_at_end_count": 253,
+  "environment_seed": 20260910,
+  "latency_seed": 271828,
+  "deterministic": true
+}

latency-aware/mikasa-intercept-grab-fast/small-policy/ppo-qwenoft-rtx3090-v1/train.yaml

@@ -0,0 +1,40 @@
+env:
+  env_fps: 20
+  obs_fps: 20
+  noop_action:
+  - 0
+  - 0
+  - 0
+  - 0
+  - 0
+  - 0
+  - 0
+executor:
+  simulated_worker_capacity: 1
+latency:
+  method: temporal
+  seed: 0
+  profile_path: profiles/qwenoft/1x-rtx3090/mikasa_intercept_grab_fast/instance_3a0d42681a03715c/profile.json
+scheduler:
+  ordering_policy: latest_ready
+  hold_policy: hold
+training:
+  num_envs: 1024
+  num_eval_envs: 16
+  num_steps: 60
+  num_minibatches: 32
+  update_epochs: 8
+  total_timesteps: 150000000
+  eval_freq: 25
+  track: true
+  exp_name: small-ppo-mikasa-intercept-grab-fast-qwenoft-rtx3090-temporal-scratch-baseline-settings
+  checkpoint: null
+  resume: null
+  seed: 123
+  num_eval_steps: 60
+  learning_rate: 0.0001
+  target_kl: 0.2
+profile_source:
+  repo_id: latency-sensitive-bench/profiles
+  revision: 66414463f2a609316da63935941a79abf0a73ce9
+  path: profiles/qwenoft/1x-rtx3090/mikasa_intercept_grab_fast/instance_3a0d42681a03715c