Festivus

Action Chunking with Transformers (ACT) policy for bimanual cube transfer. Trained in simulation on the ALOHA sim-transfer-cube-human task. This is the reference ACT checkpoint from the LeRobot project.

ACT policy for bimanual peg insertion task. The policy learns to coordinate both arms to insert a peg into a socket. Reference checkpoint from the LeRobot library.

Diffusion Policy trained on the Push-T task from Chi et al. (RSS 2023). Uses DDPM-based action generation conditioned on visual observations. The Push-T task requires pushing a T-shaped block to a target pose.

Diffusion Policy applied to the PickPlaceCan task from the Robomimic benchmark. Demonstrates state-of-the-art imitation learning on the Robomimic suite, outperforming prior BC and BC-RNN baselines.

Octo-Base: a 93M parameter generalist robot policy from the Octo team (UC Berkeley). Pre-trained on 800k episodes from the Open X-Embodiment dataset spanning 22 robot embodiments. Supports language and goal-image conditioning.