Deterministic Diffusion for Sequential Tasks

Anonymous Authors

Submitted to ICLR 2024

α-(de)blending Based

Below are short videos showing the deblending process in the instructive example experiment described in Sec. 4.2 of the paper. 

Figure 1 (Source distribution affects deblending): The target distribution is bi-modal Gaussian distribution centered at (1,1) and (3,3) and 10 steps are taken at inference time.

Figure 2 (two-stage deblending): The target distribution is bi-modal Gaussian distribution centered at (1,1) and (3.5,3.5).  Initial source distribution (red) is N((-2,-2), I)

Figure 3 (effects of conditioning): target distribution of a bi-modal Gaussian conditioned on a Bernouli random variable C: w.p. 0.5 the bi-modal Gaussian is centered around (−6, −6) and (−4, −4), or around (4, 4) and (6, 6).

Rectified-Flow Based

Below are short videos showing the rectified flow process in the same setting as the  instructive example experiment described in Sec. 4.2 of the paper, but based on the Rectified Flow algorithm instead of α-(de)blending.   All experiments are after three reflow operation.  The results show  that our method is independent of the mapping algorithm between distributions - correct selection of the initial distribution is advantageous for improved results, regardless of whether the underlying algorithm is Rectified Flow or Iterative α-(de)blending.  

Figure 1 (Source distribution affects rectified flow): The target distribution is bi-modal Gaussian distribution centered at (1,1) and (3,3) and 10 steps are taken at inference time.

Figure 2 (two-stage rectified flow): The target distribution is bi-modal Gaussian distribution centered at (1,1) and (3.5,3.5).  Initial source distribution (red) is N((-2,-2), I)

Figure 3 (effects of conditioning): target distribution of a bi-modal Gaussian conditioned on a Bernouli random variable C: w.p. 0.5 the bi-modal Gaussian is centered around (−6, −6) and (−4, −4), or around (4, 4) and (6, 6).

BAIR Videos

Below, we present 15-frame videos generated by our algorithms and different baselines conditioned on the same initial frame (for a total 16 frames).  The condition frame is outlined in red.  The 25-DDIM movement is at uneven speed.  ASPeeD FVD value is better than 25-DDPM FVD (see Sec. 5.1 of the paper).   

PHYRE Videos

Below, we present 50-frame videos generated by our algorithms and different baselines conditioned on the same 15 frames (total 65 frames).  The condition frames are outlined in red. ASPeeD FVD value is better than the baselines (see Sec. 5.1 of the paper).   

Tool-Hang Trajectories

The following videos show trajectories with the same initial conditions on the Tool-Hang task generated by ASPeeD and 10-step DDPM baseline.  In this task, a robot arm assembles a frame consisting of a base piece and hook piece by inserting the hook into the base, and hangs a wrench on the hook.  See section 5.2 of the paper for more details.

Multi-Modal Behavior

In these videos different trajectories on the Push-T task generated by ASPeeD are presented, demonstrating its multi-modal predictions.  The goal is to cover the green "T" shape mark with the gray "T" shape object.  The blue dot is the end-effector of the robot.  The red plus sign marks the next action of the robot.  See subsection 5.2.1 in the paper.