RIFT slashes robot action latency by 68% with future tokens

A team of researchers led by Chushan Zhang from the Australian National University has introduced RIFT (Rollout-free Imagination via Future Tokens), a groundbreaking approach to address a critical bottleneck in robotics: the latency introduced by iterative video rollouts in World Action Models (WAMs). WAMs condition robot actions on predicted futures, but the standard closed-loop rollout process adds significant deployment latency. The team's key insight is that while future values and their assigned positions are critical for execution, the iterative rollout process itself may not be necessary at deployment time.

The researchers demonstrate that for certain models like Joint and Cosmos-2, reusing a fixed final key/value cache preserves execution quality with minimal displacement error (1.7-1.9 cm) and high success rates (97.9%-98.2%). Building on this observation, RIFT generates a complete future key/value cache in a single backbone pass using learned anticipation tokens, eliminating the need for iterative rollouts during deployment. On the LIBERO benchmark, RIFT achieves 98.8% success—comparable to rollout-based methods like Joint (98.4%), IDM (98.5%), and LingBot-VA (98.6%)—while slashing action-chunk latency by 68.2% to 89.1%. On RoboTwin 2.0, RIFT sets new benchmarks with 92.9% success in clean scenes and 92.6% in randomized scenes.