*Real Robot Setup*

Evaluation

We trained diffusion policy to perform bottle uncorking task and evaluated its performance by varying bottle's placement across different regions on the table.

Figure 1 illustrates the spatial variations in the bottle's positions during the experiments.

                                                                                        *Figure 1: Variations in the positions of the bottle during rollouts*

Metrics:

To evaluate the performance of the model we divided the task in 3 chunks.

1. Chunk 1: One robotic arm approaches and successfully grasps the bottle.
2. Chunk 2: The second robotic arm approaches and secures a grasp on the cork.
3. Chunk 3: Robot successfully separates the cork from the bottle.