vcad-assembly - SKILL.md Agent Skill

name: vcad-assembly description: Build multi-part assemblies with joints and run physics simulations. Use when the user asks about robot arms, mechanisms, hinges, joints, physics simulation, reinforcement learning environments, or assembly of multiple parts.

vcad Assembly & Physics Simulation

Create assemblies from parts, connect them with joints, and run physics simulations via the vcad MCP tools.

Assembly Structure

An assembly is defined inside create_cad_document alongside the parts:

{
  "parts": [...],
  "assembly": {
    "instances": [...],
    "joints": [...],
    "ground": "base-1"
  }
}

Instances

Each instance places a part in the assembly:

{
  "id": "arm-1",
  "part": "arm",
  "name": "Lower Arm",
  "position": {"x": 0, "y": 0, "z": 50},
  "rotation": {"x": 0, "y": 0, "z": 0}
}

id — unique identifier (referenced by joints)
part — name of a part in the parts array
position — initial position in mm (optional)
rotation — initial rotation in degrees (optional)

Joints

Joints connect two instances and define degrees of freedom:

{
  "id": "shoulder",
  "name": "Shoulder Joint",
  "parent": "base-1",
  "child": "arm-1",
  "type": "revolute",
  "axis": "y",
  "parent_anchor": {"x": 0, "y": 0, "z": 50},
  "child_anchor": {"x": 0, "y": 0, "z": 0},
  "limits": [-90, 90],
  "state": 0
}

parent — instance ID, or null for world/ground
child — instance ID
type — joint type (see below)
axis — "x", "y", "z", or custom {x, y, z} vector
parent_anchor / child_anchor — connection points in local coordinates
limits — [min, max] (degrees for revolute, mm for slider)
state — initial angle or position

Ground

Set "ground" to the instance ID of the fixed/base part. This anchors the kinematic chain.

Joint Types

Type	DOF	Description	State Unit
`fixed`	0	Rigid attachment, no motion	n/a
`revolute`	1	Rotation around axis (hinge)	degrees
`slider`	1	Translation along axis (piston)	mm
`cylindrical`	2	Rotation + translation on same axis	degrees
`ball`	3	Omnidirectional rotation	n/a

Physics Simulation Tools

The gym-style tools simulate assembly dynamics using phyz:

Tool	Purpose
`create_robot_env`	Initialize simulation from assembly document
`gym_reset`	Reset to initial state, returns observation
`gym_step`	Apply action, advance physics, returns observation + reward + done
`gym_observe`	Get current state without stepping
`gym_close`	Clean up simulation

Simulation Workflow

create_cad_document (with assembly)
  → create_robot_env (document, end_effector_ids)
    → gym_reset (env_id)
      → gym_step (env_id, action_type, values) [loop]
        → gym_close (env_id)

create_robot_env

{
  "document": "<IR document from create_cad_document>",
  "end_effector_ids": ["gripper-1"],
  "dt": 0.004167,
  "substeps": 4,
  "max_steps": 1000
}

Returns: env_id, num_joints, action_dim, observation_dim.

gym_step

{
  "env_id": "sim_1",
  "action_type": "torque",
  "values": [0.5, -0.3]
}

Action types:

"torque" — apply torque in Nm to each joint
"position" — set target position (degrees for revolute, mm for slider)
"velocity" — set target velocity (deg/s or mm/s)

The values array must have one entry per joint.

Returns:

observation — joint positions, velocities, end effector poses
reward — scalar reward signal
done — true if episode ended (max steps or termination)

gym_reset / gym_observe

Both return the observation object:

{
  "joint_positions": [0.0, 0.0],
  "joint_velocities": [0.0, 0.0],
  "end_effector_poses": [
    {"position": {"x": 0, "y": 0, "z": 100}, "orientation": {"x": 0, "y": 0, "z": 0, "w": 1}}
  ]
}

Complete Example: 2-DOF Robot Arm

{
  "parts": [
    {
      "name": "base",
      "primitive": {"type": "cylinder", "radius": 25, "height": 10},
      "material": "steel"
    },
    {
      "name": "arm",
      "primitive": {"type": "cube", "size": {"x": 10, "y": 10, "z": 80}},
      "material": "aluminum"
    }
  ],
  "assembly": {
    "instances": [
      {"id": "base-1", "part": "base"},
      {"id": "lower-arm", "part": "arm", "position": {"x": 0, "y": 0, "z": 10}},
      {"id": "upper-arm", "part": "arm", "position": {"x": 0, "y": 0, "z": 90}}
    ],
    "joints": [
      {
        "id": "shoulder",
        "parent": "base-1",
        "child": "lower-arm",
        "type": "revolute",
        "axis": "y",
        "parent_anchor": {"x": 0, "y": 0, "z": 10},
        "child_anchor": {"x": 5, "y": 5, "z": 0},
        "limits": [-90, 90]
      },
      {
        "id": "elbow",
        "parent": "lower-arm",
        "child": "upper-arm",
        "type": "revolute",
        "axis": "y",
        "parent_anchor": {"x": 5, "y": 5, "z": 80},
        "child_anchor": {"x": 5, "y": 5, "z": 0},
        "limits": [0, 135]
      }
    ],
    "ground": "base-1"
  }
}

After creating the document, run a simulation:

create_robot_env with end_effector_ids: ["upper-arm"]
gym_reset to get initial observation
gym_step with action_type: "torque" and values per joint
Repeat step 3 for your control loop
gym_close when done

Tips

Every assembly needs a ground instance — this is the fixed reference
Joint anchors are in the instance's local coordinate frame
Revolute limits are in degrees, slider limits in mm
State outside limits is clamped automatically
Use position action type for precise pose control, torque for dynamic simulation
end_effector_ids must reference valid instance IDs in the assembly
Forward kinematics propagates transforms from ground through the joint chain