We developed RECAP Physical Intelligence to apply RL and interventions to π0.6, achieving high success rates and throughput on several challenging tasks! Watching these policies operate successfully for hours gives an appreciation for what the method can do

Interventions teach new recovery behaviors and advantage conditioning drives improvement from autonomous rollouts, turning π0.6 into π*0.6. If you watch closely you will find subtle behaviors drawn out by RECAP that are not only exciting to watch but crucial for high reliability

When picking the tasks to focus on for π*0.6 we wanted to do something that, if successful, could provide real value. For box assembly, this meant replicating @DandelionChoco's existing workflow, increasing task horizon and inheriting additional challenges

We show that on these challenging real-world tasks, more iterations continue improving the performance of π*0.6. This emphasizes the value of learning from online experience and points a direction for how we may be able to scale the deployment of robots in the real world

And finally a shoutout to an incredible team that was not only critical for bringing this together but also made it a blast to work on! If you are interested in more details you can find them in our blog post and paper:

@physical_int can I have a job on your team? will forgo salary for 6 months.

@physical_int strange how robots look productive but i just feel existential hunger