Sanctuary AI introduced TIGER, which converts action chunks produced by imitation policies into dense, task-oriented rewards for robotic manipulation in reinforcement learning. Experiments showed that TIGER reduced the number of interactions needed to reach a 50% success rate by 20% to 57% on test tasks. It recorded zero safety resets during the first 1,000 real-robot steps, compared with 35 to 38 for the baseline. On a rotating socket task with only static insertion demonstrations, TIGER achieved a 90% success rate.
Source:Physical AI for Industrial Automation · Read Original
Leave a Reply