Evaluating Lightweight Contact Feedback in Dual-Operator Collaborative Humanoid Teleoperation
In a 16-pair study where two people each teleoperated a humanoid to lift and flip a crate together, all 32 operators ranked fingertip vibration cues above sound cues, but completion time, failures and workload showed no detectable difference between vision only and any feedback mode, Georgia Tech researchers report in a preprint. Practice mattered more: the harder flip-over got about 19% faster per block and roughly halved over the session.
Does returning a humanoid's fingertip contact to its operator, by vibration, sound or both, help two operators manipulate one object together with two robots?
When two people carry an object they feel each other's moves through it. Two operators driving two humanoids through headsets see stereo video but cannot feel when their robot makes or loses contact, so they rely on vision and talking.
Headset-based humanoid teleoperation systems have mostly been single-operator without contact feedback; the closest two-operator XR humanoid system the authors found returns no contact feedback and reports no study with operator pairs.
The XRT framework maps tactile skin on each robot hand's finger patches and palm to vibrotactile cues on the operator's fingers and to spatial sounds from a Quest 3 headset's open-ear speakers. Two synchronised humanoid stations, kneeling with fixed torsos so only arms, hands and neck move, performed a crate flip-over and flip-back under four conditions in a counterbalanced within-subject design, with feedback delivery logged and verified.
All 32 operators ranked vibration above sound and 18 ranked it most effective; vibration cues were rated far easier to interpret (6.44 vs 3.03) and to localise to a finger (7.47 vs 4.19). Delivery matched the scheduled condition in 256 of 256 robot-takes, with 22,182 cue events. Completion time, failures, NASA-TLX workload, discomfort and partner awareness showed no detected condition differences; success was 120 of 128 attempts. The minimum detectable completion-time effect was about 45% at 80% power, so smaller effects remain open. Within-dyad attempt-to-attempt variation (SD x1.51) exceeded the largest condition contrast (x1.16), and only 19% of operators felt they could tell how much force they applied.
The study is underpowered for effects below about 45%, 27 of 32 operators were new to teleoperation and cybersickness limited familiarisation. Robot hands degraded during the study, one thumb lost abduction after a collision, and the skin covers only finger patches and palm, so 29% of contact time (back of hand, wrist, forearm) produced no cue. The robots did not walk or balance. It is a preprint.
Teleoperation is how most humanoid makers collect training data and run early deployments, and multi-robot tasks are next. The study warns that operator preference for haptic cues is not evidence of faster or safer work, and that practice effects can swamp interface effects in short trials. Vendors selling haptic teleoperation kits should expect buyers to ask for task-level data.
The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.