Skip to content

CURRENT_IMPLEMENTATION.md

Latest commit

 

History

History
133 lines (94 loc) · 4.42 KB

CURRENT_IMPLEMENTATION.md

File metadata and controls

133 lines (94 loc) · 4.42 KB
Raw

Current Implementation Truth

Last consolidated for main-branch cleanup.

System Story

The project is a modular L1-to-RL manipulation prototype:

Natural language command
-> Qwen / MCP semantic bridge
-> IntentPacket
-> RL skill request
-> kinematic Approach -> Finisher
-> Gazebo/RViz demo visualization and controlled execution path

The final motor-control mainline is Approach -> Finisher. Dock-Coarse, Bridge, acceptance maps, and readiness classifiers were diagnostic experiments that helped remove unnecessary modules from the final path.

L1 / L2 / L3 Contract

  • L1: Qwen/MCP resolves semantic intent, object ID, source/target slots, constraints, and skill pipeline.
  • L2: learned policy stack performs rollout and skill-level decisions.
  • L3: deterministic execution/safety layer sends bounded control to ROS2/Gazebo.

Hard boundary:

L1 must not output raw joint actions, torques, or joint trajectories.

Current Kinematic Skill Result

Official local workspace result:

Stage Success Final Pos Error Final Ori Error
0 1.00 1.67 mm 0.0106 rad
1 1.00 1.67 mm 0.0123 rad
2 1.00 1.82 mm 0.0139 rad
3 1.00 2.14 mm 0.0164 rad
4 1.00 2.53 mm 0.0165 rad
5 0.93 2.89 mm 0.0208 rad

Interpretation: the local kinematic skill works reliably inside the trained workspace. This does not imply full workspace or full tray transport.

Workspace Expansion

Later training pushed beyond Stage 5. The best home-start expansion shell reaches useful behavior through Stage 8 and partial behavior in Stage 9-11, but Stage 10/11 are still manual stress shells, not the full continuous reachable workspace.

Random-start / mixed-start coverage experiment:

Eval Split Success Mean Pos Error Mean Ori Error Main Failure
Known workspace random-start 0.802 3.87 mm 0.0185 rad position / timeout
Frontier random-start 0.240 260 mm 0.611 rad position
Full reachable stress 0.219 424 mm 1.156 rad position

Interpretation: the policy is beginning to behave like a mixed-start goal-conditioned controller in known regions, but frontier/full-reachable coverage is still partial.

Route Curriculum

The route curriculum uses dense q_goal targets as policy observations, not as direct controller commands.

Official result:

Metric Value
Baseline full483 success fraction 0.0435
Baseline longest prefix 21
Prefix120 sequential success 1.0
Prefix120 longest prefix 120
Full483 probe success fraction 0.4741
Full483 probe longest prefix 170
First failure index 171
First failure reason position

full483 success = 0.4741 is a waypoint-level/probe fraction, not a 47.41% probability of completing the whole holder1 -> holder8 route.

Gazebo / RViz Demo Status

The Gazebo demo path is for visual proof of integration:

  • natural language command is resolved into an L1 skill request,
  • target markers and route/subtask markers are published,
  • terminal shows L1 -> L2 -> L3 state transitions,
  • the arm moves in Gazebo using the demo runtime path.

Current demo claims:

  • L1 semantic contract works.
  • RViz/Gazebo visualization works for target/subtask display.
  • Local learned-policy motion can be shown.

Current demo limitations:

  • full holder1 -> holder8 tray transport is not solved,
  • object contact/friction is not validated,
  • camera view is for demonstration/context, not robust vision grounding,
  • Gazebo launch should use the original external scene launch directly when recording.

Recommended Recording Flow

Terminal 1:

cd /home/jerry/.openclaw/workspace/repos/personal/RL_brain_trainer
source /opt/ros/jazzy/setup.zsh
source external/ENPM662_Group4_FinalProject/install/setup.zsh
ros2 launch kitchen_robot_description gazebo.launch.py use_sim_time:=true headless:=false

Terminal 2:

cd /home/jerry/.openclaw/workspace/repos/personal/RL_brain_trainer
source /opt/ros/jazzy/setup.zsh
source external/ENPM662_Group4_FinalProject/install/setup.zsh
rviz2 -d config/rviz/phase3a_demo.rviz

Terminal 3:

cd /home/jerry/.openclaw/workspace/repos/personal/RL_brain_trainer
bash scripts/final/run_live_gz_screen_recording_demo.sh local_skill --no-launch-scene

Alternative planning showcase:

bash scripts/final/run_live_gz_screen_recording_demo.sh tray_like_transport --no-launch-scene