Ep#33: A Smooth Sea Never Made a Skilled SAILOR: Robust Imitation via Learning to Search

In this talk

1. 1:35

   Why imitation needs recovery

   The researchers describe how demonstration-trained policies can fail when a robot reaches situations missing from its training data. SEALER aims to improve recovery without collecting additional human demonstrations, adding planning to the learned policy.

2. 3:44

   How SEALER searches and learns

   A diffusion policy proposes action sequences, while a learned world model predicts their outcomes and a reward model scores them against expert behavior. An MPPI planner searches for residual adjustments, and refined plans can be distilled back into the policy. Training combines expert demonstrations with on-policy rollouts so the world model covers both kinds of states.

3. 13:11

   Results across tasks

   The team reports that SEALER outperformed diffusion policies on several simulated manipulation tasks, especially with fewer demonstrations. On nut assembly, for example, the diffusion policy succeeded 20% of the time with 25 demonstrations, compared with 87% for SEALER. On a door task, SEALER also solved the task with one demonstration where the diffusion policy stayed below 10%.

4. 23:40

   Planning, data, and model limits

   Reward-model visualizations showed low scores around some policy failures and higher scores at key stages such as grasping and insertion. More demonstrations, interaction data, or MPPI samples can improve results, but poor world-model or reward-model coverage can still limit performance. The team said the method had not yet been tested on a real robot.

5. 40:11

   What could come next

   For real-world use, the researchers discussed stronger pretrained world models and visual representations, lower planning latency, and distilling planner improvements into the base policy. They also raised safety measures for online rollouts and adaptive compute allocation, using reward scores and uncertainty to decide when more search may help.