FlashDexRetarget cuts dexterous retargeting compute

The important idea is amortizing retargeting across an entire motion collection instead of optimizing each demonstration independently, though the compute comparison also changes the underlying RL algorithm.

• –Achieves 90% success on 50 motions using about 30 GPU-hours, versus CHORD’s 46% success at roughly 3,000 GPU-hours.
• –Combines object point clouds, hand-object distance features, future trajectory encoding, and separate left/right actor-critic networks.
• –Scaling tests cover collections of up to 1,000 motions, suggesting the approach becomes more valuable as demonstration datasets grow.
• –Real-robot replays show wiping, pouring, and lid-closing, but broader deployment still depends on clean demonstrations, accurate geometry, and simulation assumptions.
• –The headline advantage should be interpreted cautiously because FlashDexRetarget uses FlashSAC while the compared CHORD implementation uses PPO. ([analysis](https://pith.science/paper/2610.01849))