Add reduced UR10e deploy-reach slice on MLX/mac-sim

Author: AIFlowML

.github/workflows/mlx-macos.yml

@@ -245,6 +245,7 @@ jobs:
         bin_stack_cfg = parse_env_cfg("Isaac-Stack-Cube-Bin-Franka-IK-Rel-Mimic-v0", device="cpu", num_envs=4)
         stack_instance_cfg = parse_env_cfg("Isaac-Stack-Cube-Instance-Randomize-Franka-v0", device="cpu", num_envs=4)
         teddy_cfg = parse_env_cfg("Isaac-Lift-Teddy-Bear-Franka-IK-Abs-v0", device="cpu", num_envs=4)
+        ur10e_cfg = parse_env_cfg("Isaac-Deploy-Reach-UR10e-v0", device="cpu", num_envs=4)
         payload = evaluate_mlx_task(
             "h1-rough",
             num_envs=4,
@@ -260,6 +261,7 @@ jobs:
             "bin_stack_cfg_type": type(bin_stack_cfg).__name__,
             "stack_instance_cfg_type": type(stack_instance_cfg).__name__,
             "teddy_cfg_type": type(teddy_cfg).__name__,
+            "ur10e_cfg_type": type(ur10e_cfg).__name__,
             "trainable_tasks": list(list_trainable_mlx_tasks()),
             "eval_task": payload["task"],
             "episodes_completed": payload["episodes_completed"],
@@ -335,6 +337,7 @@ jobs:
           source/isaaclab/test/backends/test_mac_anymal_c.py \
           source/isaaclab/test/backends/test_mac_anymal_c_rough.py \
           source/isaaclab/test/backends/test_mac_franka_reach.py \
+          source/isaaclab/test/backends/test_mac_ur10e_deploy_reach.py \
           source/isaaclab/test/backends/test_mac_franka_lift.py \
           source/isaaclab/test/backends/test_mac_franka_teddy_bear_lift.py \
           source/isaaclab/test/backends/test_mac_franka_stack_instance_randomize.py \
@@ -432,6 +435,21 @@ jobs:
           --num-envs 8 --updates 1 --rollout-steps 8 --epochs-per-update 1 --episode-length-s 0.5 \
           --checkpoint logs/mlx/franka_reach_policy.npz --eval-interval 1
 
+    - name: Run MLX UR10e deploy-reach smoke
+      run: |
+        PYTHONPATH=.:source/isaaclab .venv/bin/python \
+          scripts/reinforcement_learning/mlx/evaluate_task.py \
+          --task ur10e-deploy-reach \
+          --num-envs 8 --episodes 1 --episode-length-s 0.5 --max-steps 512 --no-random-actions
+
+    - name: Run MLX UR10e deploy-reach training smoke
+      run: |
+        PYTHONPATH=.:source/isaaclab .venv/bin/python \
+          scripts/reinforcement_learning/mlx/train_task.py \
+          --task ur10e-deploy-reach \
+          --num-envs 8 --updates 1 --rollout-steps 8 --epochs-per-update 1 --episode-length-s 0.5 \
+          --checkpoint logs/mlx/ur10e_deploy_reach_policy.npz --eval-interval 1
+
     - name: Run MLX Franka lift smoke
       run: |
         PYTHONPATH=.:source/isaaclab .venv/bin/python \
@@ -565,6 +583,8 @@ jobs:
         assert "cartpole-rgb-camera" in list_mlx_tasks()
         assert "cartpole-depth-camera" in list_mlx_tasks()
         assert "franka-reach" in list_mlx_tasks()
+        assert "ur10e-deploy-reach" in list_mlx_tasks()
+        assert "ur10e-deploy-reach" in list_trainable_mlx_tasks()
         assert "franka-lift" in list_mlx_tasks()
         assert "franka-teddy-bear-lift" in list_mlx_tasks()
         assert "franka-stack-instance-randomize" in list_mlx_tasks()
@@ -574,19 +594,19 @@ jobs:
         assert "franka-cabinet" in list_mlx_tasks()
         assert "franka-open-drawer" in list_mlx_tasks()
         train_payload = train_mlx_task(
-            "franka-bin-stack",
+            "ur10e-deploy-reach",
             num_envs=8,
             updates=1,
             rollout_steps=8,
             epochs_per_update=1,
             hidden_dim=32,
-            checkpoint="logs/mlx/franka_bin_stack_wrapper_policy.npz",
+            checkpoint="logs/mlx/ur10e_deploy_reach_wrapper_policy.npz",
             eval_interval=1,
             episode_length_s=0.5,
             seed=31,
         )
         payload = evaluate_mlx_task(
-            "franka-bin-stack",
+            "ur10e-deploy-reach",
             checkpoint=train_payload["checkpoint_path"],
             episodes=1,
             episode_length_s=0.5,
@@ -731,10 +751,10 @@ jobs:
         assert len(payload["benchmarks"]) >= 5
         print(payload["tasks"])
         dashboard = json.loads(Path("logs/benchmarks/mlx/smoke-dashboard.json").read_text(encoding="utf-8"))
-        assert dashboard["summary"]["rollout_task_count"] == 16
+        assert dashboard["summary"]["rollout_task_count"] == 17
         assert dashboard["summary"]["training_task_count"] == 0
         trend = json.loads(Path("logs/benchmarks/mlx/smoke-trend.json").read_text(encoding="utf-8"))
-        assert trend["summary"]["task_count"] == 16
+        assert trend["summary"]["task_count"] == 17
         planner_payload = json.loads(Path("logs/planner/mac-planner-smoke.json").read_text(encoding="utf-8"))
         assert planner_payload["planner"]["implementation"] == "joint-space-linear-interpolation"
         assert planner_payload["plan"]["waypoint_count"] == 6
@@ -751,8 +771,8 @@ jobs:
         assert stereo_payload["depth_mm_mean"] is not None
         assert Path(stereo_payload["left_rgb_path"]).is_file()
         runtime_payload = json.loads(Path("logs/runtime/mac-runtime-diagnostics.json").read_text(encoding="utf-8"))
-        assert runtime_payload["runtime"]["supported_tasks"]["public_task_count"] == 18
-        assert runtime_payload["sim"]["supported_tasks"]["current_mac_native_count"] == 16
+        assert runtime_payload["runtime"]["supported_tasks"]["public_task_count"] == 19
+        assert runtime_payload["sim"]["supported_tasks"]["current_mac_native_count"] == 17
         sensor_payload = json.loads(Path("logs/benchmarks/mlx/sensor-smoke.json").read_text(encoding="utf-8"))
         assert sensor_payload["task_group"] == "sensor-mac-native"
         assert sensor_payload["cpu_fallback_detected"] is False
@@ -778,15 +798,17 @@ jobs:
         assert release_payload["bin_stack_cfg_type"] == "MacFrankaBinStackEnvCfg"
         assert release_payload["stack_instance_cfg_type"] == "MacFrankaStackInstanceRandomizeEnvCfg"
         assert release_payload["teddy_cfg_type"] == "MacFrankaTeddyBearLiftEnvCfg"
+        assert release_payload["ur10e_cfg_type"] == "MacUR10eDeployReachEnvCfg"
         assert "anymal-c-rough" in release_payload["trainable_tasks"]
         assert "h1-rough" in release_payload["trainable_tasks"]
+        assert "ur10e-deploy-reach" in release_payload["trainable_tasks"]
         assert "franka-teddy-bear-lift" in release_payload["trainable_tasks"]
         assert "franka-stack-instance-randomize" in release_payload["trainable_tasks"]
         assert "franka-bin-stack" in release_payload["trainable_tasks"]
         assert release_payload["eval_task"] == "h1-rough"
         assert release_payload["episodes_completed"] == 1
         release_runtime_payload = json.loads(Path("logs/runtime/release-runtime-diagnostics.json").read_text(encoding="utf-8"))
-        assert release_runtime_payload["runtime"]["supported_tasks"]["public_task_count"] == 18
+        assert release_runtime_payload["runtime"]["supported_tasks"]["public_task_count"] == 19
         release_eval_payload = json.loads(Path("logs/runtime/release-eval.json").read_text(encoding="utf-8"))
         assert release_eval_payload["task"] == "h1-rough"
         assert release_eval_payload["episodes_completed"] == 1
@@ -824,7 +846,7 @@ jobs:
         full_payload = json.loads(Path("logs/benchmarks/mlx/full-smoke.json").read_text(encoding="utf-8"))
         assert full_payload["task_group"] == "full"
         full_dashboard = json.loads(Path("logs/benchmarks/mlx/full-smoke-dashboard.json").read_text(encoding="utf-8"))
-        assert full_dashboard["summary"]["rollout_task_count"] == 20
+        assert full_dashboard["summary"]["rollout_task_count"] == 21
         assert full_dashboard["summary"]["training_task_count"] == 1
         full_trend = json.loads(Path("logs/benchmarks/mlx/full-smoke-trend.json").read_text(encoding="utf-8"))
         assert any(entry["kind"] == "training" for entry in full_trend["tasks"])

PORT_TO_MLX_TODO.md

@@ -102,6 +102,7 @@ without pausing for replanning after every small success.
 - `DONE` Eighth trainable Franka manipulation slice landed for `Isaac-Stack-Cube-Instance-Randomize-Franka-v0` by keeping the shared analytic stack substrate but adding explicit variant-id observations, deterministic distinct support/movable object sampling, public MLX wrapper/CLI support, benchmark coverage, semantic baseline refresh, and focused backend tests instead of dishonestly aliasing it to plain stack
 - `DONE` Ninth trainable Franka manipulation slice landed for `Isaac-Stack-Cube-Bin-Franka-IK-Rel-Mimic-v0` by keeping the shared three-cube stack substrate, pinning the support cube to an explicit bin anchor, exposing that anchor through the observation and benchmark contracts, refreshing the semantic baseline, and validating replay/train plus wrapper/CLI coverage
 - `DONE` The `Isaac-Stack-Cube-Bin-Franka-IK-Rel-Mimic-v0` alias now carries explicit reduced-contract metadata across the lazy registry, parsed env config, public MLX task spec, installed CLI payloads, and backend diagnostics so the mac path does not silently imply upstream mimic/imitation parity
+- `DONE` Tenth trainable manipulation slice landed for `Isaac-Deploy-Reach-UR10e-v0` with a reduced analytic pose-tracking backend, honest reduced-contract metadata (`semantic_contract="reduced-analytic-pose"`), public MLX wrapper/CLI support, benchmark coverage, semantic baseline refresh, and direct replay/train smoke coverage; the upstream `Isaac-Deploy-Reach-UR10e-Play-v0` alias now resolves to the same canonical mac-native slice
 
 ## Phase A: Import And Packaging Safety
 
@@ -754,7 +755,7 @@ without pausing for replanning after every small success.
 This queue exists so work can continue without waiting for a new plan. The documented v1 board above is now closed for the current public MLX/mac slice, so the next queue is follow-on parity work:
 
 - Hardware validation is now done for the backend-local stereo path against live ZED 2i capture through a camera-authorized Terminal host plus `zed-sdk-mlx`; retained host-local probe artifacts include `/tmp/isaaclab-zed-probe-live-final.json` and `/tmp/isaaclab-zed-probe-live-final.yuv`.
-- Port the next manipulation milestone beyond the current nine trainable Franka slices, likely a richer cabinet/drawer variant or the next multi-object manipulation workflow.
+- Port the next genuinely new manipulation milestone beyond the current ten trainable manipulation slices, likely a richer non-Franka cabinet/drawer variant or the next multi-object workflow that cannot be represented honestly as an alias of the existing reduced slices.
 - Replace the next remaining locomotion or contact/support `mx.compile` helper with a true custom Metal kernel only after the root-step tranche proves benchmark-positive and semantically stable.
 - Grow the planner/ROS prototypes carefully: richer process/message interoperability layers around the new world-state and joint-trajectory envelopes while still avoiding CUDA/NITROS assumptions.
 - Keep the generic runtime metadata honest: only advertise generic sensor/runtime capabilities that are actually exposed through backend-neutral APIs, and push task-specific or tooling-only support into explicit diagnostic fields instead of broad parity flags.
@@ -768,6 +769,7 @@ PYTHONPATH=.:source/isaaclab:source/isaaclab_rl .venv/bin/pytest \
   scripts/tools/test/test_bootstrap_isaac_sources.py \
   scripts/tools/test/test_bootstrap_uv_mlx.py \
   source/isaaclab_rl/test/test_import_safety.py \
+  source/isaaclab_rl/test/test_mlx_cli.py \
   source/isaaclab_rl/test/test_mlx_wrapper.py \
   source/isaaclab/test/backends/test_runtime.py \
   source/isaaclab/test/backends/test_task_registry.py \
@@ -790,9 +792,14 @@ PYTHONPATH=.:source/isaaclab:source/isaaclab_rl .venv/bin/pytest \
   source/isaaclab/test/backends/test_mac_anymal_c.py \
   source/isaaclab/test/backends/test_mac_anymal_c_rough.py \
   source/isaaclab/test/backends/test_mac_franka_reach.py \
+  source/isaaclab/test/backends/test_mac_ur10e_deploy_reach.py \
   source/isaaclab/test/backends/test_mac_franka_lift.py \
+  source/isaaclab/test/backends/test_mac_franka_teddy_bear_lift.py \
+  source/isaaclab/test/backends/test_mac_franka_stack_instance_randomize.py \
   source/isaaclab/test/backends/test_mac_franka_bin_stack.py \
   source/isaaclab/test/backends/test_mac_franka_stack.py \
+  source/isaaclab/test/backends/test_mac_franka_stack_rgb.py \
+  source/isaaclab/test/backends/test_mac_franka_cabinet.py \
   source/isaaclab/test/backends/test_mac_franka_open_drawer.py \
   source/isaaclab/test/backends/test_mac_h1.py -q
 ```