Back to the archive

The Motion Forge / Hybrid controller · study ongoing

Reinforcement learning for robotic pick-and-place

A hybrid UR10e controller combining a learned PPO approach with deterministic grasping, Lula transport, and verified basket placement in Isaac Lab.

PythonPyTorchNVIDIA Isaac LabIsaac SimPPOLula
* CHECK
SUBJECT
6DOF Robot
TYPE
Robotics + deep RL
STATUS
Hybrid controller · study ongoing

* 144 / 256 stable holds in the frozen baseline.

PROOF SIGNAL

144 / 256stable holds in the frozen baseline

The mission

What the system does

A hybrid UR10e controller combining a learned PPO approach with deterministic grasping, Lula transport, and verified basket placement in Isaac Lab.

  • 34-dimensional structured observations and target-conditioned PPO approach
  • Deterministic contact, recovery, release, and two-second success verification
  • Lula transport reached 20/48 verified placements versus 5/48 for the legacy route
  • Four randomization profiles, three training seeds, and reproducible artifact validation

Engineering log

Decisions behind the build

What the system handles, and why it works this way.

BOSS 01

Sparse success

A full pick-and-place reward arrives too late for early learning. Staged objectives provide useful intermediate signals.

BOSS 02

Simulation stability

Contact-heavy manipulation is sensitive to setup. Collision handling and controlled curriculum changes keep training interpretable.

System map

How signals move

A plain technical view. The map still works without hovering.

01
ApproachPPO / PyTorch

Target-conditioned positioning

02
ContactState machine

Grasp, lift, hold, and retry

03
TransportLula

Clearance and basket alignment

04
VerificationIsaac Lab

Release and stable placement

LOG ENTRY / WHAT I LEARNED

“Reward design is an interface between an engineering goal and what an agent can actually discover.”

Source signal

Inspect the work.