FastWAM LIBERO ROBOTWIN sim packages

.gitignore

@@ -77,3 +77,9 @@ dist/
 
 # docs
 docs/_build/
+
+# Per-package local working artifacts (raw rollouts/plots/profiles kept locally for
+# observation; never committed - curated small clips go in each package's assets/)
+packages/**/artifacts/
+# Repo-root holding for artifacts pending manual package assignment (never committed)
+artifacts_unsorted/

README.md

@@ -22,9 +22,10 @@ These dockerfiles will also be pushed and actively maintained in their original
 | LLM                     | [`ollama`](packages/llm/ollama), [`llamacpp`](packages/llm/llamacpp), [`lmstudio`](packages/llm/lmstudio), [`GEPA`](packages/llm/gepa), [`Lemonade`](packages/llm/lemonade-sdk)   |
 | VLM                     | [`Gemma3`](packages/vlm/gemma3), [`SmolVLM`](packages/vlm/smolvlm), [`Phi-4`](packages/vlm/phi4), [`LFM2-VL`](packages/vlm/lfm2vl) |
 | VLA                     | [`OpenVLA`](packages/vla/openvla), [`SmolVLA`](packages/vla/smolvla), [`GR00T-N1.5`](packages/vla/gr00t), [`openpi`](packages/vla/openpi), [`CogACT`](packages/vla/cogact), [`MolmoAct`](packages/vla/molmoact), [`MolmoAct2`](packages/vla/molmoact2) |
+| WAM     | [`FastWAM`](packages/wam/fastwam) |
 | Graphics                     | [`O3DE`](packages/graphics/o3de) |
 | Robotics                | [`ROS 2`](packages/ros/ros), [`Gazebo`](packages/ros/gazebo), [`LeRobot`](packages/robotics/lerobot), [`ACT`](packages/robotics/act), [`RAI`](packages/robotics/rai), [`CaP-X`](packages/robotics/capx)    |
-| Simulation                |  [`Genesis`](packages/robotics/genesis), [`PyDrake`](packages/robotics/pydrake)  |
+| Simulation                |  [`Genesis`](packages/robotics/genesis), [`PyDrake`](packages/robotics/pydrake), [`LIBERO`](packages/simulation/libero), [`RoboTwin`](packages/simulation/robotwin)  |
 | Vision                  | [`OpenCV`](packages/vision/opencv), [`SAM`](packages/vision/sam), [`MobileSAM`](packages/vision/mobilesam), [`ncnn`](packages/vision/ncnn), [`DINOv3`](packages/vision/dinov3), [`SAM3`](packages/vision/sam3), [`Ultralytics`](packages/vision/ultralytics) |
 | Ryzen AI NPU                |  [`XDNA`](packages/npu/xdna), [`IRON`](packages/npu/iron), [`NPUEval`](packages/npu/npueval), [`Ryzen AI CVML`](packages/npu/ryzenai_cvml)  |
 | Adaptive SoCs           | [`PYNQ.remote`](packages/adaptive-socs/pynq-remote) |

packages/simulation/libero/Dockerfile

@@ -0,0 +1,52 @@
+# Copyright(C) 2026 Advanced Micro Devices, Inc. All rights reserved.
+# SPDX-License-Identifier: MIT
+#
+# Model-agnostic LIBERO sim base for Strix Halo (gfx1151): LIBERO/MuJoCo/robosuite
+# closed-loop stack (headless EGL) + sim_libero harness. Policies chain on top and
+# are selected at runtime via POLICY_FACTORY; weights/datasets are never baked here.
+
+ARG BASE_IMAGE
+FROM ${BASE_IMAGE}
+
+ENV DEBIAN_FRONTEND=noninteractive \
+    HF_HUB_DISABLE_TELEMETRY=1 \
+    TOKENIZERS_PARALLELISM=false
+
+WORKDIR /ryzers
+
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+      git cmake build-essential \
+      libegl-dev libgles-dev libosmesa6-dev libglew-dev libglu1-mesa libglu1-mesa-dev && \
+    rm -rf /var/lib/apt/lists/*
+
+# numpy pinned 1.26.4 (robosuite numba has no py3.12/numpy-2 wheel); constraint keeps the
+# base ROCm torch+numpy. LIBERO installs --no-deps; missing runtime imports added explicitly.
+# libero ships no top-level __init__.py, so PYTHONPATH resolves it.
+ENV PYTHONPATH=/opt/LIBERO:/opt/sim
+RUN pip install --no-cache-dir "numpy==1.26.4" && \
+    pip freeze | grep -iE "^(torch|torchvision|torchcodec|numpy)==" > /opt/torch-constraints.txt && \
+    PIP_CONSTRAINT=/opt/torch-constraints.txt pip install --no-cache-dir \
+      "robosuite==1.4.0" "bddl==1.0.1" "mujoco==3.3.2" easydict thop \
+      "future==1.0.0" "cloudpickle==3.1.2" "gym==0.25.2" termcolor \
+      "imageio>=2.34" "imageio-ffmpeg>=0.5" "pillow>=10" && \
+    git clone --depth 1 https://github.com/Lifelong-Robot-Learning/LIBERO /opt/LIBERO && \
+    PIP_CONSTRAINT=/opt/torch-constraints.txt pip install --no-cache-dir --no-deps -e /opt/LIBERO
+
+# The sim harness (Policy seam + runners + RandomPolicy).
+COPY lib/ /opt/sim/
+
+# torch.load compat shim as a startup .pth so it applies to every Python process
+# (incl. a policy's own eval script). See _torch_compat.
+RUN cp /opt/sim/sim_libero_compat.pth "$(python -c 'import sysconfig; print(sysconfig.get_paths()["purelib"])')/"
+
+# First `import libero.libero` prompts to write ~/.libero/config.yaml; feed defaults so it
+# bakes in and never prompts at runtime. Also verifies the stack imports (no GPU render at
+# build time; the sanity demo exercises rendering at run time).
+RUN printf 'N\nN\nN\nN\nN\n' | python -c "import mujoco, robosuite; from libero.libero import benchmark; from libero.libero.envs import OffScreenRenderEnv; import sim_libero; from sim_libero.policy import load_policy; print('OK mujoco', mujoco.__version__, '| sim_libero + LIBERO envs import cleanly')"
+
+COPY test.py /ryzers/
+COPY demos/ /ryzers/demos/
+RUN chmod +x /ryzers/demos/*.sh
+
+CMD ["python", "/ryzers/test.py"]

packages/simulation/libero/README.md

@@ -0,0 +1,31 @@
+### LIBERO
+
+This package provides the LIBERO simulation base (MuJoCo/EGL) as a Ryzer on AMD Ryzen AI
+Max+ 395 (Strix Halo, gfx1151) under ROCm 7.14. It exposes a model-agnostic `Policy` seam
+that WAM and VLA packages chain on for closed-loop and interactive rollouts.
+
+### Build
+
+```sh
+ryzers build libero
+ryzers run     # test.py: sim import + headless render sign-of-life
+```
+
+### Example: LIBERO with FastWAM
+
+```sh
+ryzers build libero fastwam --name fastwam-libero
+ryzers run --name fastwam-libero /ryzers/demos/demo_closedloop_libero.sh
+```
+
+<p align="center">
+  <img src="assets/libero_fastwam.gif" width="480">
+  <br><em>FastWAM closed-loop rollout in LIBERO.</em>
+</p>
+
+### References
+
+- Upstream: https://github.com/Lifelong-Robot-Learning/LIBERO (the Dockerfile shallow-clones the default branch)
+
+Copyright (C) 2026 Advanced Micro Devices, Inc. All rights reserved.
+SPDX-License-Identifier: MIT

packages/simulation/libero/config.yaml

@@ -0,0 +1,34 @@
+# Copyright(C) 2026 Advanced Micro Devices, Inc. All rights reserved.
+# SPDX-License-Identifier: MIT
+
+# ROCm 7.14 base (rocm/pytorch torch 2.10.0).
+init_image: rocm/pytorch:rocm7.14_ubuntu24.04_py3.12_pytorch_release_2.10.0
+# Model-agnostic sim base; chain a policy: ryzers build libero <policy>
+
+gpu_support: true
+x11_display: false
+
+environment_variables:
+- "HSA_OVERRIDE_GFX_VERSION=11.5.1"
+- "MUJOCO_GL=egl"
+- "PYOPENGL_PLATFORM=egl"
+- "TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1"
+- "TORCH_BLAS_PREFER_HIPBLASLT=0"
+# Distinct mount target so this base's outputs don't collide with a chained policy's /outputs.
+- "OUT_DIR=/sim_outputs"
+# Sim/harness knobs (override from the host: VAR=... ryzers run /ryzers/demos/demo_*.sh).
+- 'SUITE=${SUITE:-}'
+- 'TASK_ID=${TASK_ID:-}'
+- 'SEED=${SEED:-}'
+- 'STEPS=${STEPS:-}'
+- 'PORT=${PORT:-}'
+- 'VIEW_RES=${VIEW_RES:-}'
+- 'VIDEO_RES=${VIDEO_RES:-}'
+- 'RENDER_RES=${RENDER_RES:-}'
+- 'MAX_STEPS=${MAX_STEPS:-}'
+- 'RT_HZ=${RT_HZ:-}'
+# Policy selection: module:function returning a sim_libero.Policy (default RandomPolicy).
+- 'POLICY_FACTORY=${POLICY_FACTORY:-}'
+
+volume_mappings:
+- "$PWD/workspace/simulation-libero/outputs:/sim_outputs"

packages/simulation/libero/demos/demo_interactive.sh

@@ -0,0 +1,15 @@
+#!/usr/bin/env bash
+# Copyright(C) 2026 Advanced Micro Devices, Inc. All rights reserved.
+# SPDX-License-Identifier: MIT
+#
+# Interactive LIBERO demo (chunk-replay) over HTTP/MJPEG. Default RandomPolicy; set
+# POLICY_FACTORY for a real model. View at http://localhost:PORT (ssh -L PORT:localhost:PORT <host>).
+#   ryzers run /ryzers/demos/demo_interactive.sh
+#   POLICY_FACTORY=fastwam_libero_policy:build_policy ryzers run /ryzers/demos/demo_interactive.sh
+set -euo pipefail
+export SUITE="${SUITE:-libero_object}"
+export TASK_ID="${TASK_ID:-0}"
+export SEED="${SEED:-1000}"
+export PORT="${PORT:-8080}"
+export OUT_DIR="${OUT_DIR:-/sim_outputs}"
+exec python -m sim_libero.interactive_server

packages/simulation/libero/demos/demo_interactive_rt.sh

@@ -0,0 +1,17 @@
+#!/usr/bin/env bash
+# Copyright(C) 2026 Advanced Micro Devices, Inc. All rights reserved.
+# SPDX-License-Identifier: MIT
+#
+# Real-time LIBERO demo: sim steps at wall-clock RT_HZ and the robot HOLDs while the
+# policy plans, so latency is visible. Default RandomPolicy; set POLICY_FACTORY for a model.
+# View at http://localhost:PORT (ssh -L PORT:localhost:PORT <host>).
+#   ryzers run /ryzers/demos/demo_interactive_rt.sh
+#   POLICY_FACTORY=fastwam_libero_policy:build_policy ryzers run /ryzers/demos/demo_interactive_rt.sh
+set -euo pipefail
+export SUITE="${SUITE:-libero_object}"
+export TASK_ID="${TASK_ID:-0}"
+export SEED="${SEED:-1000}"
+export PORT="${PORT:-8081}"
+export RT_HZ="${RT_HZ:-20}"
+export OUT_DIR="${OUT_DIR:-/sim_outputs}"
+exec python -m sim_libero.interactive_server_rt

packages/simulation/libero/demos/demo_sim_sanity.sh

@@ -0,0 +1,15 @@
+#!/usr/bin/env bash
+# Copyright(C) 2026 Advanced Micro Devices, Inc. All rights reserved.
+# SPDX-License-Identifier: MIT
+#
+# Model-free sanity rollout: rolls the built-in RandomPolicy through one scene on the
+# iGPU (headless EGL) and saves an MP4. Exercises the ROCm/EGL render + MuJoCo + encode path.
+#   ryzers run /ryzers/demos/demo_sim_sanity.sh
+#   SUITE=libero_goal TASK_ID=2 STEPS=120 ryzers run /ryzers/demos/demo_sim_sanity.sh
+set -euo pipefail
+export SUITE="${SUITE:-libero_object}"
+export TASK_ID="${TASK_ID:-0}"
+export SEED="${SEED:-1000}"
+export STEPS="${STEPS:-80}"
+export OUT_DIR="${OUT_DIR:-/sim_outputs}"
+exec python -m sim_libero.sanity

packages/simulation/libero/examples/template_policy.py

@@ -0,0 +1,68 @@
+# Copyright(C) 2026 Advanced Micro Devices, Inc. All rights reserved.
+# SPDX-License-Identifier: MIT
+"""COPY-ME template: plug your own model into the LIBERO sim base.
+
+The sim base is model-agnostic; you build FROM it, ship an adapter like this, and select
+it at runtime (do NOT edit this package).
+
+Steps
+-----
+1. Chain on the sim base in your model's Dockerfile, installing your model UNDER the
+   base's torch+numpy pins (needs numpy 1.26.4) so the LIBERO/robosuite/MuJoCo stack is
+   not broken (FastWAM's scripts/strip_cuda_torch.py is a working reference):
+
+       ARG BASE_IMAGE
+       FROM ${BASE_IMAGE}
+       COPY scripts/adapters/ /opt/model-adapters/
+
+   Build:   ryzers build libero <yourmodel> --name <yourmodel>-libero
+
+2. Implement the two methods below (load in build_policy, obs -> action chunk in
+   predict_action_chunk).
+
+3. Run with your factory on PYTHONPATH:
+
+       PYTHONPATH=/opt/model-adapters:$PYTHONPATH \
+       POLICY_FACTORY=template_policy:build_policy \
+         ryzers run --name <yourmodel>-libero /ryzers/demos/demo_interactive.sh
+
+Contract
+--------
+obs           : raw LIBERO/robosuite observation dict. Useful keys:
+                  obs["agentview_image"]            HxWx3 uint8 (3rd-person; flipped)
+                  obs["robot0_eye_in_hand_image"]   HxWx3 uint8 (wrist; flipped)
+                  obs["robot0_eef_pos"]             (3,)  end-effector position
+                  obs["robot0_eef_quat"]            (4,)  end-effector orientation (xyzw)
+                  obs["robot0_gripper_qpos"]        (2,)  gripper joint positions
+instruction   : str, the natural-language task.
+return        : np.ndarray [T, 7] float32 = (dx,dy,dz,droll,dpitch,dyaw, gripper), LIBERO
+                OSC_POSE delta control; gripper {-1 open, +1 close}. Harness executes the
+                first `replan_steps` rows, then calls you again.
+"""
+import numpy as np
+
+from sim_libero.policy import Policy
+
+
+class TemplatePolicy(Policy):
+    name = "template"
+    replan_steps = 5      # env steps executed per predicted chunk before replanning
+    num_steps_wait = 5    # no-op settle steps at episode start
+
+    def __init__(self, model):
+        self.model = model
+
+    def reset(self, instruction):
+        # Called once per episode; clear per-episode caches here.
+        pass
+
+    def predict_action_chunk(self, obs, instruction):
+        # TODO: preprocess obs, run inference, return actions in the OSC_POSE delta space above.
+        raise NotImplementedError("wire your model here")
+        # return np.zeros((self.replan_steps, 7), dtype=np.float32)
+
+
+def build_policy():
+    # TODO: load your checkpoint / processor once here (env knobs: CKPT, DATASET_STATS, ...).
+    model = None
+    return TemplatePolicy(model)

packages/simulation/libero/lib/sim_libero/__init__.py

@@ -0,0 +1,14 @@
+# Copyright(C) 2026 Advanced Micro Devices, Inc. All rights reserved.
+# SPDX-License-Identifier: MIT
+"""Model-agnostic LIBERO simulator harness.
+
+LIBERO env glue, a `Policy` seam, closed-loop/interactive runners, and a RandomPolicy.
+Policies plug in via a `build_policy() -> Policy` factory selected by POLICY_FACTORY.
+"""
+from sim_libero._torch_compat import patch_torch_load
+
+patch_torch_load()
+
+from sim_libero.policy import Policy, load_policy  # noqa: E402
+
+__all__ = ["Policy", "load_policy"]

packages/simulation/libero/lib/sim_libero/_torch_compat.py

@@ -0,0 +1,25 @@
+# Copyright(C) 2026 Advanced Micro Devices, Inc. All rights reserved.
+# SPDX-License-Identifier: MIT
+"""torch.load compatibility shim for LIBERO's pickled assets.
+
+PyTorch >= 2.6 defaults torch.load to weights_only=True, rejecting LIBERO's pickled
+init-states. Restore weights_only=False (trusted assets shipped in the image). Applied via
+a startup .pth (covers any policy's eval script) and from sim_libero.__init__.
+"""
+
+
+def patch_torch_load():
+    try:
+        import torch
+    except ImportError:
+        return
+    if getattr(torch.load, "_sim_libero_patched", False):
+        return
+    _orig = torch.load
+
+    def _load(*args, **kwargs):
+        kwargs.setdefault("weights_only", False)
+        return _orig(*args, **kwargs)
+
+    _load._sim_libero_patched = True
+    torch.load = _load

packages/simulation/libero/lib/sim_libero/envutil.py

@@ -0,0 +1,21 @@
+# Copyright(C) 2026 Advanced Micro Devices, Inc. All rights reserved.
+# SPDX-License-Identifier: MIT
+"""Env-var readers that treat an empty string as unset.
+
+ryzers run passes optional knobs as `-e VAR=${VAR:-}` (empty when unset), so plain
+os.environ.get would return "" and break int()/float(); these fall back to the default.
+"""
+import os
+
+
+def env_str(key, default):
+    val = os.environ.get(key)
+    return val if val not in (None, "") else default
+
+
+def env_int(key, default):
+    return int(env_str(key, str(default)))
+
+
+def env_float(key, default):
+    return float(env_str(key, str(default)))