Sanctuary AI's TIGER turns imitation policies into a reward signal, cutting real-robot RL safety resets to zero

Sanctuary AI said in a research post on Oct 2 that TIGER (Task-Space Imitation Guidance for Efficient Reinforcement Learning) uses an imitation policy's predicted action chunk not as commands but as a short-horizon end-effector target, giving a reinforcement learner a dense reward at every control step instead of a single success bit at the end.

The company reports 20% to 57% fewer steps to reach 50% success on every sparse-reward manipulation task it tested, zero safety resets in the first 1,000 real-robot steps against 35 to 38 for the baseline, and 90% insertion into a rotating socket when demonstrations only showed a stationary one.

The results are the company's own and not peer reviewed. MEKAWIRE records Sanctuary's last disclosed priced round as a $58.5 million Series A in March 2022, followed by strategic investments with undisclosed amounts in 2024 and May 2026.