LieVLA trained models (private)
Trained models of the Lie-VLA study (code and configs: github.com/Grik301/LieVLA; results: 40 unseen + 40 seen scenarios).
| folder | contents |
|---|---|
| lievla/checkpoints/uav | Lie-VLA, drone, 2007 demonstrations (matched comparison) |
| lievla/checkpoints_v3/long, dagger_fixed, dagger_badlabels | long training (80k steps), long + DAgger, DAgger without 4-DoF projection (failed) |
| lievla/checkpoints_v4/dagger1, dagger2 | DAgger and two DAgger rounds |
| lievla/checkpoints_v5/servo, servo_dagger | Lie-VLA + SE(3) grasp head (grasp servo), drone |
| lievla/checkpoints_v6/residual | residual-RL (CEM) servo correction parameters |
| lievla/checkpoints_f1000/franka, checkpoints_f1000_servo/franka | Lie-VLA, Franka 1000 demonstrations, without / with grasp head |
| lievla/checkpoints_dart/, checkpoints_dart_servo/ | Lie-VLA retrained with DART data, without / with grasp head (franka, uav) |
| lievla/checkpoints_diff/* | SE(3) diffusion head (v3: per-step scale + EDM preconditioning) |
| lievla/study1/ | first studies: Franka 363-demo Lie-VLA, drone study 1, single-task prototype (policy_se3, policy_euler, world_model) |
| baselines/drone/{pi0,pi05,pi0fast,smolvla} | LeRobot baselines, drone, 2007 demonstrations (final step 3000) |
| baselines/franka_1000/{pi0,pi05,pi0fast,smolvla} | LeRobot baselines, Franka, 1000 demonstrations (final step 3000) |
A lievla.pt holds the policy state, normalisation statistics and cfg (and the grasp head under "grasp" for servo models); load it with lievla.runners2.LieVLARunner. Baseline folders are LeRobot pretrained_model directories. DART-retrained and Franka-363 baselines are stored in a separate organization repository.