When Will You Hand It Over? A Hidden Start Delay Makes Self-Play Handover Robust to Unfamiliar Partners

TL;DR A simulated Unitree G1 that learns to receive a parcel by PPO self-play succeeds in 90-95% of handovers with its own partner but only 31% with partners it has never met, because it learns when its partner starts. Adding a random hidden start delay to the partner raises success with unseen partners to 80% in both training seeds.
What they do
Cooperative self-play can teach a humanoid to take an object from a giver without any human data, but the trained pair may settle on a private convention. In a simulated parcel handover between a fixed-base Unitree G1 and a learned giver, the authors find that this convention is about when the giver starts. They fix it with one hidden rule on the giver.
How it works
- Self-play. In Isaac Lab, the G1 receives a 0.5 kg parcel from a learned partner hand that sets its reach speed, small offer corrections and release. Both act from state, not images, and take turns being updated by PPO, with a pool of past partners.
- Success measure. A supported transfer means the robot alone holds the parcel with a formed grip for 15 steps, with limits over the whole episode on unsupported steps, slip in the grasp and contact force between the robot and the partner’s hand.
- Hidden start delay (HSD). Each episode draws a 0-1 s delay during which the partner’s hand must stay still. The receiver’s observations, rewards and network are unchanged, so it must read from the partner’s motion when the offer begins.
Results
- Self-play receivers (two seeds) succeed 95.0 / 89.5% of the time with their own partner but 30.5 / 30.5% with scripted partners. Delaying their own partner by 1 s drops success from 95.0 to 6.0% and from 87.0 to 33.5%.
- After 300 more updates with HSD, success with unseen scripted partners reaches 80.5 / 80.0%, against 35.0 / 41.0% for a control given the same extra training. Across eleven held-out partner conditions the mean rises from 30.5 / 27.2% to 78.8 / 62.7%.
- Episodes with robot-partner contact above 20 N fall from 10.7 / 7.9% to 1.9 / 2.5%. Keeping the partner frozen gives the same robustness.
Limitations
Simulation only, with a simulated partner hand, state inputs instead of images and two training seeds. The scripted partners are designed behaviours, and their start times overlap the training delays. Every receiver stays below 5% with a scripted partner that withdraws if the parcel moves more than 3 cm before release.
Figures




Abstract
Cooperative self-play can teach a humanoid to receive an object without any human data, but the partner it meets after training is never the one it trained with; we show that a fluent self-play handover policy breaks with unfamiliar partners, identify the main cause, and repair it. In a physics-based parcel handover between a Unitree G1 and a learned giver, both trained by alternating PPO self-play, the receiver completes 95% and 90% of handovers with its own partner (two training seeds) but only 31% with scripted partners it has never met. Paired interventions isolate the cause: under a shared reward the learned giver always starts its offer at once, so the receiver learns a single rendezvous, and delaying its own partner by one second cuts success from 95% to 6% (87% to 34% in the second seed). We introduce hidden start-delay randomisation (HSD), a one-line rule on the learned partner that the receiver never observes. After 300 further updates, success with unseen scripted partners rises from 31% to 80% in both seeds, against 35--41% for a budget-matched control. The gain holds on a fresh scenario seed, on eleven held-out partner conditions and on partners that start later than any training delay; it cuts high-force contact three- to five-fold and needs no retraining of the partner.
Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.
