DreamMachines/cube_pnp_d0_act_r1

cube_pnp/d0_act_r1

ACT (Action Chunking with Transformers) — imitation learning that predicts short action chunks instead of single steps. See the paper.

Trained with LeDream (a fork of LeRobot).

Training

• Dataset: DreamMachines/cube_pnp_view_d0
• Step (this revision): 80000
• Batch size: 32
• Total steps target: 80000
• Optimizer: adamw (lr 3e-05)

Available revisions

Every training checkpoint is uploaded as its own git revision; the latest is also on main. Load a specific step with:

from lerobot.policies.factory import make_policy
from lerobot.configs.policies import PreTrainedConfig

cfg = PreTrainedConfig.from_pretrained("DreamMachines/cube_pnp_d0_act_r1", revision="v002000")
policy = make_policy(cfg)

Available: v002000, v004000, v006000, v008000, v010000, v012000, v014000, v016000, v018000, v020000, v022000, v024000, v026000, v028000, v030000, v032000, v034000, v036000, v038000, v040000, v042000, v044000, v046000, v048000, v050000, v052000, v054000, v056000, v058000, v060000, v062000, v064000, v066000, v068000, v070000, v072000, v074000, v076000, v078000, v080000

Quick eval on a real DK1

make eval MODEL=DreamMachines/cube_pnp_d0_act_r1

See docs/source/policy_pi05_README.md for π₀.₅-specific notes.