Fix to ros2 action list

.github/build-config.json

@@ -7,5 +7,5 @@
     {"name": "gfx120x", "pytorch_whl": "gfx120X-all", "rocm_sdk": "gfx120x"}
   ],
   "default_gpu_target": "gfx1151",
-  "courses": ["CV", "DL", "LLM", "PhySim"]
+  "courses": ["CV", "DL", "LLM", "PhySim", "Finetuning", "LocalInference"]
 }

.gitignore

@@ -388,6 +388,7 @@ vite.config.ts.timestamp-*
 # Project specific
 dockerfiles/Courses/DL/data/FashionMNIST/raw/
 *.vsix
+projects/LocalInference/sweep_results/
 
 # Local config overrides
 runtime/values.local.yaml

auplc_installer/catalog.py

@@ -51,6 +51,27 @@ class Course:
     Course("Course-DL", "auplc-dl", True, "dl", "Deep Learning Course"),
     Course("Course-LLM", "auplc-llm", True, "llm", "Large Language Model Course"),
     Course("Course-PhySim", "auplc-physim", True, "physim", "Physical Simulation Course"),
+    Course(
+        "Course-Finetuning",
+        "auplc-finetuning",
+        True,
+        "finetuning",
+        "Fine-tuning on GPUs: from cloud to robot",
+    ),
+    Course(
+        "Course-LocalInference",
+        "auplc-localinference",
+        True,
+        "local-inference",
+        "Local inference of embodied AI",
+    ),
+    Course(
+        "Course-RLLearning",
+        "auplc-rl-learning",
+        True,
+        "rl-learning",
+        "Reinforcement learning for robotics",
+    ),
 )
 
 COURSE_KEYS_ALL: tuple[str, ...] = tuple(c.key for c in COURSE_CATALOG)
@@ -81,7 +102,16 @@ class Course:
 
 BASE_TEAM_MAPPING: dict[str, list[str]] = {
     "cpu": ["cpu", "code-cpu"],
-    "gpu": ["code-gpu", "Course-CV", "Course-DL", "Course-LLM", "Course-PhySim"],
+    "gpu": [
+        "code-gpu",
+        "Course-CV",
+        "Course-DL",
+        "Course-LLM",
+        "Course-PhySim",
+        "Course-Finetuning",
+        "Course-LocalInference",
+        "Course-RLLearning",
+    ],
     "official": [
         "cpu",
         "gpu",
@@ -91,15 +121,29 @@ class Course:
         "Course-DL",
         "Course-LLM",
         "Course-PhySim",
+        "Course-Finetuning",
+        "Course-LocalInference",
+        "Course-RLLearning",
+    ],
+    "AUP": [
+        "Course-CV",
+        "Course-DL",
+        "Course-LLM",
+        "Course-PhySim",
+        "Course-Finetuning",
+        "Course-LocalInference",
+        "Course-RLLearning",
     ],
-    "AUP": ["Course-CV", "Course-DL", "Course-LLM", "Course-PhySim"],
     "native-users": [
         "code-cpu",
         "code-gpu",
         "Course-CV",
         "Course-DL",
         "Course-LLM",
         "Course-PhySim",
+        "Course-Finetuning",
+        "Course-LocalInference",
+        "Course-RLLearning",
         "cpu",
         "gpu",
     ],
@@ -112,6 +156,9 @@ class Course:
         "Course-DL",
         "Course-LLM",
         "Course-PhySim",
+        "Course-Finetuning",
+        "Course-LocalInference",
+        "Course-RLLearning",
     ],
 }
 

auplc_installer/overlay.py

@@ -36,6 +36,9 @@
     "Course-DL": "auplc-dl",
     "Course-LLM": "auplc-llm",
     "Course-PhySim": "auplc-physim",
+    "Course-Finetuning": "auplc-finetuning",
+    "Course-LocalInference": "auplc-localinference",
+    "Course-RLLearning": "auplc-rl-learning",
 }
 
 

dockerfiles/Courses/Finetuning/build.sh

@@ -0,0 +1,43 @@
+#!/usr/bin/env bash
+# Copyright (C) 2026 Advanced Micro Devices, Inc. All rights reserved.
+# SPDX-License-Identifier: MIT
+
+set -euo pipefail
+
+# Hardcoded workshop asset bundle location: drop mm2_workshop_assets.zip into projects/Finetuning/
+# (next to the notebooks) and it gets baked into the image. Nothing to configure.
+PROJECT_DIR="../../../projects/Finetuning"
+ASSETS_ZIP="${PROJECT_DIR}/mm2_workshop_assets.zip"
+
+WORK_ASSETS=""
+cleanup() {
+  rm -rf course_data
+  [ -n "${WORK_ASSETS:-}" ] && rm -rf "${WORK_ASSETS}"
+  return 0
+}
+trap cleanup EXIT
+
+# Stage the course notebooks + helper scripts into the build context, EXCLUDING the large asset
+# bundle so it is never baked into /ryzers/notebooks.
+rm -rf course_data
+mkdir -p course_data
+cp -r "$PROJECT_DIR"/. course_data/
+rm -rf course_data/mm2_workshop_assets course_data/mm2_workshop_assets.zip
+
+# Bake the workshop assets if the bundle is present; otherwise build a code-only image (CI/registry).
+BUILD_EXTRA=()
+if [ -f "${ASSETS_ZIP}" ]; then
+  echo "unpacking mm2_workshop_assets.zip (one-time; the bundle is large, this can take a while)..."
+  # Unpack OUTSIDE the docker build context (extracting here would bloat the context sent to the
+  # daemon). /var/tmp is disk-backed and typically large enough for the ~16 GB bundle.
+  WORK_ASSETS="$(mktemp -d "${TMPDIR:-/var/tmp}/auplc-ft-assets.XXXXXX")"
+  unzip -q "${ASSETS_ZIP}" -d "${WORK_ASSETS}"
+  echo "baking workshop assets into the image from: ${WORK_ASSETS}/mm2_workshop_assets"
+  BUILD_EXTRA+=(--build-context "assets=${WORK_ASSETS}/mm2_workshop_assets" --build-arg "FINAL=with-assets")
+else
+  echo "no mm2_workshop_assets.zip in projects/Finetuning/ -> building code-only image (assets NOT baked in)."
+fi
+
+DOCKER_BUILDKIT=1 docker build ${BASE_IMAGE:+--build-arg BASE_IMAGE="$BASE_IMAGE"} \
+  ${BUILD_EXTRA[@]+"${BUILD_EXTRA[@]}"} \
+  -t ghcr.io/amdresearch/auplc-finetuning:latest .

dockerfiles/Courses/Finetuning/fastwam/README.md

@@ -0,0 +1,123 @@
+### FastWAM
+
+This package runs [FastWAM](https://github.com/yuantianyuan01/FastWAM) — a Wan2.2-TI2V-5B
+world-action model (T5 text encoder + Wan VAE + video/action DiT) — on AMD Ryzen AI Max+
+395 (Strix Halo, `gfx1151`) under ROCm 7.2.2. Direct PyTorch port: upstream code runs on
+the base image's ROCm torch; only the CUDA torch pins are stripped.
+
+It is a **slim policy/model layer that ships no simulator**, and the reference consumer of
+the simulator packages' `Policy` seam. It composes on:
+
+- the plain ROCm base → non-sim demos (smoke / latency / open-loop / videogen);
+- the `simulation/libero` base → closed-loop + interactive LIBERO;
+- the `simulation/robotwin` base → closed-loop + interactive RoboTwin 2.0.
+
+The same policy layer composes on all three because the FastWAM install is pinned to the
+base image's torch + numpy (so a plain base's numpy 2.x and a sim base's numpy 1.26.4 both
+work). Weights/datasets are fetched by the scripts below; sim assets come from the sim base.
+
+### Build
+
+```sh
+# Standalone (non-sim demos + model sign-of-life):
+ryzers build fastwam --name fastwam
+ryzers run --name fastwam                 # test.py: ROCm torch + GPU + deps sign-of-life
+
+# Chain on a simulator base for closed-loop / interactive rollouts:
+ryzers build libero   fastwam --name fastwam-libero
+ryzers build robotwin fastwam --name fastwam-robotwin
+```
+
+Artifacts are written to `workspace/*/outputs`. For faster/gated HF downloads set
+`HF_TOKEN`. The ~12 GB Wan2.2 base is fetched automatically on the first model run.
+
+```sh
+ryzers run --name fastwam /ryzers/scripts/download_checkpoints.sh    # LIBERO + RoboTwin ckpts
+ryzers run --name fastwam /ryzers/scripts/download_datasets.sh       # open-loop / video data
+```
+
+The chain drives a sim base's model-agnostic `Policy` seam via a runtime adapter
+(`adapters/fastwam_{libero,robotwin}_policy.py`, selected by `POLICY_FACTORY`); these
+adapters double as the worked reference for wiring any VLA/WAM into the sim bases (see each
+`simulation/*` README). The RoboTwin closed-loop instead runs RoboTwin's own
+`script/eval_policy.py` against `experiments/robotwin/fastwam_policy`
+(`EVALUATION.robotwin_root=/opt/RoboTwin`).
+
+### Demos
+
+| Demo | Base | What it does |
+|---|---|---|
+| `demos/demo_smoke.sh` | plain | Load checkpoint, one `infer_action`; cold/steady latency + VRAM. |
+| `demos/demo_latency.sh` | plain | Per-part latency (T5 / VAE / world prefill / plan) + SDPA backends. |
+| `demos/demo_openloop.sh` | plain | Replay GT observations, overlay predicted vs GT action chunks + MAE. |
+| `demos/demo_videogen.sh` | plain | Imagine future frames from the first observation; GT-vs-imagined clips. |
+| `demos/demo_closedloop_libero.sh` | `libero` | Closed-loop LIBERO rollouts (MuJoCo/EGL) + success rate. |
+| `demos/demo_interactive_libero.sh` / `_rt.sh` | `libero` | Interactive LIBERO over HTTP/MJPEG. |
+| `demos/demo_closedloop_robotwin.sh` | `robotwin` | Closed-loop RoboTwin 2.0 rollouts (SAPIEN Vulkan RT) + success rate. |
+| `demos/demo_interactive_robotwin.sh` / `_rt.sh` | `robotwin` | Interactive RoboTwin over HTTP/MJPEG. |
+
+```sh
+ryzers run --name fastwam        /ryzers/demos/demo_smoke.sh
+ryzers run --name fastwam-libero /ryzers/demos/demo_closedloop_libero.sh
+TASKS="click_bell lift_pot" NUM_EPISODES=10 \
+  ryzers run --name fastwam-robotwin /ryzers/demos/demo_closedloop_robotwin.sh
+ryzers run --name fastwam-robotwin /ryzers/demos/demo_interactive_robotwin.sh   # http://localhost:8082
+```
+
+### Open-loop replay
+
+Predicted action chunks track ground truth over 100 episodes: mean normalized MAE
+**0.0222** (LIBERO) / **0.0208** (RoboTwin), action inference ~1.5 s.
+
+<p align="center">
+  <img src="assets/d1_libero_per_dim_mae.png" alt="open-loop per-dim MAE, LIBERO" width="700">
+  <br><em>Per-dimension normalized MAE (LIBERO).</em>
+</p>
+<p align="center">
+  <img src="assets/d1_libero_ep00.png" alt="open-loop GT-vs-pred overlay, LIBERO episode 0" width="700">
+  <br><em>GT (solid) vs predicted (dashed) action chunks, LIBERO episode 0.</em>
+</p>
+
+### Video imagination
+
+Joint path imagines the future video + actions (GT left, imagined right). Steady-state
+joint latency ~18.6 s (LIBERO) / ~21.9 s (RoboTwin) for a 33-frame clip at 20 denoise
+steps (the first call pays a one-time ROCm warmup).
+
+<p align="center">
+  <img src="assets/d2_libero.gif" alt="GT vs imagined, LIBERO" width="600">
+  <br><em>Ground truth vs imagined future (LIBERO).</em>
+</p>
+<p align="center">
+  <img src="assets/d2_robotwin.gif" alt="GT vs imagined, RoboTwin" width="600">
+  <br><em>Ground truth vs imagined future (RoboTwin).</em>
+</p>
+
+### Closed-loop LIBERO
+
+`libero_object` suite, 10 tasks × 20 trials: **199/200 (99.5%)** success, rendered headless
+via EGL. With `VISUALIZE_FUTURE=true` the slow path also renders the model's imagined future
+alongside the real rollout (GT left, imagined right; PSNR ~27.3 dB).
+
+<p align="center">
+  <img src="assets/d3_libero.gif" alt="closed-loop LIBERO rollout" width="400">
+  <img src="assets/d4_slow.gif" alt="closed-loop slow path, GT vs imagined" width="400">
+  <br><em>Closed-loop rollout (left) and slow-path GT-vs-imagined (right).</em>
+</p>
+
+### Useful knobs
+
+- Non-sim: `DATASET=libero|robotwin` (open-loop/videogen/latency), `NUM_STEPS`, `SEED`.
+- Closed-loop LIBERO: `SUITE`, `NUM_TASKS`, `NUM_TRIALS`, `VISUALIZE_FUTURE`.
+- Closed-loop RoboTwin: `TASKS`, `TASK_CONFIG`, `NUM_EPISODES`.
+- Interactive: `PORT`, `CKPT`, `DATASET_STATS`, `REPLAN_STEPS`, `NUM_INFERENCE_STEPS`.
+- `HF_TOKEN` for faster/gated downloads.
+
+### References
+
+- Upstream: https://github.com/yuantianyuan01/FastWAM (pinned in `docs/UPSTREAM_PIN.commit.txt`)
+- Model: https://huggingface.co/yuanty/fastwam
+- Datasets: https://huggingface.co/datasets/yuanty/LIBERO-fastwam · https://huggingface.co/datasets/yuanty/robotwin2.0-fastwam
+
+Copyright (C) 2026 Advanced Micro Devices, Inc. All rights reserved.
+SPDX-License-Identifier: MIT

dockerfiles/Courses/Finetuning/fastwam/adapters/fastwam_libero_policy.py

@@ -0,0 +1,115 @@
+# Copyright(C) 2026 Advanced Micro Devices, Inc. All rights reserved.
+# SPDX-License-Identifier: MIT
+"""FastWAM LIBERO policy adapter for the simulation/libero harness.
+
+Implements the model-agnostic `sim_libero.Policy` seam by wrapping the FastWAM world-action
+model. Reuses the *validated* eval machinery from experiments/libero/eval_libero_single.py
+verbatim (config compose, model instantiate + checkpoint load, processor/normalizer, and
+`_predict_action_chunk`) so interactive/closed-loop rollouts match the shipped numbers.
+
+Selected at runtime by the sim harness via
+  POLICY_FACTORY=fastwam_libero_policy:build_policy
+
+Env: CKPT, DATASET_STATS, MIXED_PRECISION (bf16), SUITE, NUM_INFERENCE_STEPS,
+REPLAN_STEPS, NUM_STEPS_WAIT, FASTWAM_REPO (/repos/fastwam).
+Requires /repos/fastwam and its experiments/libero dir on PYTHONPATH (the demo sets this).
+"""
+import os
+
+import numpy as np
+from hydra import compose, initialize_config_dir
+from hydra.core.global_hydra import GlobalHydra
+from hydra.utils import instantiate
+
+import experiments.libero.eval_libero_single as E
+from fastwam.datasets.lerobot.utils.normalizer import load_dataset_stats_from_json
+from sim_libero.policy import Policy
+
+FASTWAM_REPO = os.environ.get("FASTWAM_REPO", "/repos/fastwam")
+CONFIG_DIR = os.path.join(FASTWAM_REPO, "configs")
+DEFAULT_CKPT = "/models/fastwam_release/libero_uncond_2cam224.pt"
+DEFAULT_STATS = "/models/fastwam_release/libero_uncond_2cam224_dataset_stats.json"
+
+
+class FastwamLiberoPolicy(Policy):
+    name = "fastwam"
+
+    def __init__(self, model, processor, cfg, action_horizon, input_w, input_h, device):
+        self.model = model
+        self.processor = processor
+        self.cfg = cfg
+        self.action_horizon = action_horizon
+        self.input_w = input_w
+        self.input_h = input_h
+        self.device = device
+        self.replan_steps = int(cfg.EVALUATION.get("replan_steps", 5))
+        self.num_steps_wait = int(cfg.EVALUATION.get("num_steps_wait", 5))
+
+    def predict_action_chunk(self, obs, instruction):
+        action, _imgs, _pred = E._predict_action_chunk(
+            obs=obs,
+            task_description=instruction,
+            model=self.model,
+            processor=self.processor,
+            cfg=self.cfg,
+            action_horizon=self.action_horizon,
+            input_w=self.input_w,
+            input_h=self.input_h,
+            model_device=self.device,
+        )
+        return np.asarray(action, dtype=np.float32)
+
+
+def build_policy():
+    # ryzers passes optional knobs as empty strings; treat "" as unset.
+    ckpt = os.environ.get("CKPT") or DEFAULT_CKPT
+    stats = os.environ.get("DATASET_STATS") or DEFAULT_STATS
+    mixed = os.environ.get("MIXED_PRECISION") or "bf16"
+    suite = os.environ.get("SUITE") or "libero_object"
+
+    overrides = [
+        f"ckpt={ckpt}",
+        "gpu_id=0",
+        f"mixed_precision={mixed}",
+        f"EVALUATION.task_suite_name={suite}",
+        "EVALUATION.task_id=0",
+        "EVALUATION.num_trials=1",
+        f"EVALUATION.dataset_stats_path={stats}",
+        "EVALUATION.output_dir=/tmp/fastwam_interactive",
+    ]
+    if os.environ.get("NUM_INFERENCE_STEPS"):
+        overrides.append(f"EVALUATION.num_inference_steps={os.environ['NUM_INFERENCE_STEPS']}")
+    if os.environ.get("REPLAN_STEPS"):
+        overrides.append(f"EVALUATION.replan_steps={os.environ['REPLAN_STEPS']}")
+    if os.environ.get("NUM_STEPS_WAIT"):
+        overrides.append(f"EVALUATION.num_steps_wait={os.environ['NUM_STEPS_WAIT']}")
+
+    if GlobalHydra.instance().is_initialized():
+        GlobalHydra.instance().clear()
+    with initialize_config_dir(config_dir=CONFIG_DIR, version_base="1.3"):
+        cfg = compose(config_name="sim_libero", overrides=overrides)
+
+    device = E._resolve_eval_device(cfg)
+    dtype = E._mixed_precision_to_model_dtype(mixed)
+    model = instantiate(cfg.model, model_dtype=dtype, device=device)
+    E._load_model_checkpoint(model, str(cfg.ckpt))
+    model = model.to(device).eval()
+
+    stats_path = E._resolve_dataset_stats_path(cfg)
+    dataset_stats = load_dataset_stats_from_json(str(stats_path))
+    processor = instantiate(cfg.data.train.processor).eval()
+    processor.set_normalizer_from_stats(dataset_stats)
+
+    action_horizon_cfg = cfg.EVALUATION.get("action_horizon", None)
+    if action_horizon_cfg is None:
+        action_horizon = int(cfg.data.train.num_frames) - 1
+    else:
+        action_horizon = int(action_horizon_cfg)
+
+    video_size = cfg.data.train.get("video_size", [224, 224])
+    input_h = int(video_size[0])
+    input_w = int(video_size[1])
+
+    print(f"[fastwam_libero_policy] model ready (ckpt={ckpt}, horizon={action_horizon}, "
+          f"input={input_w}x{input_h}, replan={cfg.EVALUATION.get('replan_steps', 5)})", flush=True)
+    return FastwamLiberoPolicy(model, processor, cfg, action_horizon, input_w, input_h, device)

dockerfiles/Courses/Finetuning/fastwam/docs/UPSTREAM_PIN.commit.txt

@@ -0,0 +1,2 @@
+yuantianyuan01/FastWAM
+45d8e1458921d83f8ad6cf9ce993d371208dabd0