Docs / CAD / DCC Reinforcement Learning

CAD / DCC Reinforcement Learning

Mock-first RL for Creo, SolidWorks, Blender, and Cinema 4D, using MathNN DQN via RLAgentNet.

Host	First scenarios	Knobs (p0,p1,p2)
creo	`box_fit`, `mass_target`	LENGTH, WIDTH, HEIGHT proxies
sw	`mass_target`, `box_fit`	same parametric MDP (mock COM)
blender	`game_prop_export`, `turntable`	decimate %, subdiv, camera yaw
c4d	`cloner_grid`, `turntable`	cloner counts X/Y/Z

Safety

Mock metrics always (no license required).
Optional Creo param mirror uses bridge dry-run when started.
Live toolkit/COM writes are not enabled by this module.
Prefer body/creative dry-run freezes for any future live promotion.

Commands

cad rl status|doctor|hosts|scenarios|selftest|smoke|help
cad rl env reset <host> <scenario>
cad rl step [auto|p0+|p0-|p1+|p1-|p2+|p2-|commit|noop]
cad rl train <host> <episodes> [scenario]
cad rl eval <host> <episodes> [scenario]
cad rl prefer good|bad
cad rl checkpoint save|load [path]

creo rl train 50 box_fit
blender rl train 30 game_prop_export
c4d rl train 30 cloner_grid
sw rl train 30 mass_target
creative rl smoke

Quick start

cad rl doctor
cad rl env reset creo box_fit
cad rl step auto
cad rl train creo 50 box_fit
cad rl eval creo 10 box_fit
cad rl smoke

MDP summary

Item	Value
State dim	16 (knobs, mass/poly/bbox, interfere/fail/export, aesthetic, progress, host)
Actions	8: `p0± p1± p2± commit noop`
Learner	`RLAgentNet` DQN per host agent id `cad_dcc_<host>`
Reward	scenario-specific (envelope MSE, mass cap, poly budget, aesthetic, cloner packing)

Lua

cad_rl_reset("creo", "box_fit")
cad_rl_train("creo", 50, "box_fit")
cad_rl_eval("blender", 10, "game_prop_export")
cad_rl_prefer("good")
cad_rl_smoke()
cad_rl_selftest()

Files

CadDccRL.hpp / CadDccRL.cpp
Wired in SpeechCommands.cpp, AgentRuntime.cpp

Generated from the project markdown docs on 2026-07-24. This is a static, self-contained site.