ForceTwin

“ForceTwin achieved 87.3% goal completion across nine object-embodiment pairs, versus 59.7% for VLM-prior twins and 56.9% for kinematics-only twins”

“A human with a force-sensing handheld gripper can measure an object's true dynamics in minutes, producing a digital twin that nearly doubles robot manipulation success”

“we replace that randomization with the identified dynamics, leaving the policy architecture and training procedure unchanged”

“the wood-door policy 'succeeds on the wood door in all five trials'”

“kinematics-only twins (pure feedback) had the worst tracking errors on most objects — e.g., 91.9% tracking RMSE on the oven with Spot vs. 33.3% for ForceTwin”

“81.2% completion vs. 1.6% (VLM prior) and 1.5% (kinematics-only) on the metal door”

“Tool trajectories come from 'Project Aria MPS' (Section IV-A)”

“Ground-truth articulation axes were annotated in 'a Leica RTC360 laser scan'”

“Capture uses 'the handheld Hoi! gripper [10]'”