feat: nero simple pick devpack

ApiTests/ApiTests.zip

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8624a70dacfa32e7ee6a4c5ea1f7996f782e7629c72a8bf88f152f51a95a78c7
-size 73110815
+oid sha256:17a8254b0f2d496e9ffaae83d842650193c6330edcf90183a3e68c60ab24bfd8
+size 73116576

ApiTests/metadata.yaml

@@ -1,6 +1,6 @@
 Name: "API Tests"
 Type: Project
-Version: "1.7"
+Version: "1.8"
 EngineVersion:
   Min: "2026.2"
   Behaviour: Hide
@@ -31,6 +31,9 @@ Description: >
   range-calibrated reflectivity, and pass-through secondary returns off the wall behind.
   run_all_tests.py sweeps every scene and reports pass/fail, installing requirements.txt into
   its test interpreter first.
+  1.8 replaces the embedded Nero with the Nero pack's own 1.3 files (arm turned to face +X, nero_-prefixed
+  joint and actuator names, gripper_joint1 actuator, retuned gains, reaching home keyframe, box finger pads)
+  and renames the Nero agents' joints to match; all four Nero scenes pass against it.
 Tags:
   - tests
   - determinism

Nero/Nero.zip

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:04df0e234b18e184ae01f4115b840cbb01a4afee37512aa3417f92144fc33c62
-size 14412942
+oid sha256:237e95ec68e605f89c4148c327434880b00988c1cfae5a530577333b4e9056ff
+size 14413103

Nero/metadata.yaml

@@ -1,6 +1,6 @@
 Name: "Nero"
 Category: Robots
-Version: "1.2"
+Version: "1.3"
 Description: >
   A detailed robotic arm asset pack for the AgileX Nero 7-DoF manipulator with parallel
   gripper, including OBJ visual and STL collision assets for individual links, a grasp_site
@@ -18,6 +18,9 @@ Description: >
   joint-to-actuator lookup requires, and gives the finger pads hard, high-priority contact
   parameters so a gripped object does not creep down the pads under the engine's soft-friction
   MuJoCo settings (see README.md).
+  Pack 1.3 gives nero_arm.xml the same nero_ joint and actuator prefix as nero.xml, so neither
+  file can match the engine's Piper-specific overrides on any build, and the ApiTests project (1.8)
+  now embeds these exact files, so the Nero its four scenes test is the Nero this pack ships.
 Tags:
   - robot
   - manipulator

SimplePick/SimplePick.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b07611cb81374b73408d574b23b60f5b57567a92b4947ed23bbcd1d1852df45a
+size 31724173

SimplePick/metadata.yaml

@@ -0,0 +1,61 @@
+Name: "SimplePick"
+Type: Project
+Version: "1.0"
+EngineVersion:
+  Min: "2026.2"
+  Behaviour: Hide
+Category: Examples
+Description: >
+  A scripted pick-and-place data generator and policy-evaluation rig, built twice: once for the
+  AgileX Piper (6-DoF) and once for the AgileX Nero (7-DoF). A block spawns at a random position
+  and yaw inside a fan in front of the arm, a bin spawns elsewhere in it, and an engine-side C#
+  controller drives the arm through the scene's LimbIK motion graph to pick the block up and drop
+  it in. Every episode is recorded to a LeRobot v3.0 dataset (per-joint state, velocity and action
+  columns, end-effector and object poses, wrist and front cameras at 256x256, 50 Hz) and judged by
+  the same success test the policy rig uses, so generator and evaluator always agree.
+  Four scenes. Main (Piper) and Main-Nero record data. PolicyEval and PolicyEval-Nero run the
+  identical task with the arm driven by a trained policy over the External Action Gate instead of
+  by the script, on a seeded spawn sequence so two policies meet identical layouts;
+  PolicyEval-Nero-Holdout is the same on a second seed for confirming a winner. Tools/ holds the
+  session-to-LeRobot exporter (streams, detects the joint list, hardlinks video), the evaluation
+  client and runner (which read the robot off a checkpoint's action width and pick the scene), and
+  a training-strategy comparison harness.
+  The Nero port is the interesting half and its findings are written up in the project README and
+  in NeroArm.cs: the arm is turned to face +X because its elbow bends 123 degrees one way and 58
+  the other; the jaws slide along the grasp site's X axis where the Piper's slide along Y, a 90
+  degree difference in the commanded wrist yaw; LimbIK's Tracking Mode must be on or a redundant
+  7-DoF arm comes back on a different branch and the graph lerps every joint there; a vertical
+  tool is only reachable outside r = 0.31 m and below a height that falls with radius; and the
+  gripper actuator is named after its joint, because the engine resolves an agent's JointSpec
+  entry to an actuator by exact joint name and silently falls back to actuator 0 on a miss.
+  NOT INCLUDED, deliberately: the trained policies (198 MB each) and the recorded datasets (a
+  single 80k-episode session is 64 GB). Tools/lerobot_export.py plus lerobot-train regenerate a
+  policy from a fresh recording; the project README carries the exact commands and the measured
+  results (Piper 27/35, Nero 46/49 on 49 seeded layouts).
+  ONE THING TO FIX PER MACHINE: ObserverConfig.ColumnSchemaPath is an absolute path stored in the
+  recording scenes and will point at someone else's drive. The engine resolves it with a bare
+  filesystem check and refuses to record if it is missing, loudly, so it cannot corrupt a dataset -
+  but repoint it before recording. Evaluation does not record and needs no change.
+Tags:
+  - example
+  - robotics
+  - pick-and-place
+  - data-generation
+  - imitation-learning
+  - lerobot
+  - policy-evaluation
+  - external-action-gate
+  - piper
+  - nero
+  - limbik
+Files:
+  - 'SimplePick.hproj'
+  - 'README.md'
+  - 'Assets/Scenes/'
+  - 'Assets/Schemas/'
+  - 'Assets/Scripts/Client/Source/'
+  - 'Assets/ContentVault/'
+  - 'Assets/Materials/'
+  - 'Assets/Meshes/'
+  - 'Assets/EnvMap/'
+  - 'Tools/'

index.json

@@ -30,6 +30,7 @@
         "UnitreeG1FiveFingersHrobot",
         "UnitreeG1ThreeFingersHrobot",
         "CoffeeMachine",
-        "Nero"
+        "Nero",
+        "SimplePick"
     ]
 }