Learning Multi-Humanoid Pickup and Transport via Decentralized Object-Centric Control
Decentralized Multi-Humanoid Transport (decMHT)
Under Review
decMHT is a decentralized object-centric control abstraction that lets teams of humanoid robots cooperatively pick up, transport, and hand off shared objects using gripperless bimanual pinching, spanning single-robot pickup, multi-robot transport, and robot-to-robot handover under one policy, without inter-robot communication.
Abstract
We study cooperative multi-humanoid pickup and transport of objects with varying size, weight, and geometry, requiring robot teams of different sizes. Our approach uses decentralized object-centric control, where each humanoid is assigned a local attachment region on the shared object and learns to realize pickup and transport through gripperless bimanual pinching. This attachment-based interface provides a common control abstraction spanning single-robot pickup, cooperative multi-robot transport, and robot-to-robot handover, without per-task redesign. We find that policies trained only on single-robot pickup already transfer nontrivially to cooperative settings, suggesting that this abstraction captures much of the structure needed for coordination. At the same time, explicit multi-robot training further improves performance, showing that shared-object coupling introduces coordination dynamics that are beneficial to learn directly. We validate the approach in simulation across varying team sizes and object geometries, and demonstrate sim-to-real transfer on hardware, where the learned controllers enable real humanoids to perform cooperative manipulation tasks.
Decentralized Multi-Humanoid Transport (decMHT)
We train decMHT in large-scale parallel simulation and evaluate it on multi-robot teams, diverse payloads, and real Digit V3 humanoid hardware.
Training (Simulation)
decMHT is trained in IsaacLab with large-scale parallel simulation, where humanoids learn to pinch, lift, and transport a shared payload through the object-centric attachment interface.
Generalization to 2-10 Robots (Simulation)
A single trained controller supports teams of up to ten humanoids transporting shared payloads, without retraining for team size.
Single-Robot Pickup and Transport (Simulation)
A single humanoid learns to approach, pinch, and lift a payload using gripperless bimanual contact.
Generalization to Diverse Payloads (Simulation)
The same policy generalizes to out-of-distribution objects, including cabinets, concrete blocks, fridges, logs, and beds, with different geometries and attachment layouts.
Robot-to-Robot Handover (Simulation)
Humanoids hand off the payload to one another across multiple elevation differences and under dynamic motion, all using the same object-centric abstraction.
Single-Robot Pickup and Transport (Hardware)
A Digit V3 humanoid executes the learned pickup policy on real hardware.
Multi-Robot Pickup and Transport (Hardware)
Two Digit V3 humanoids cooperatively transport shared payloads of different sizes using the same decentralized policy.
Robot-to-Robot Handover (Hardware)
Humanoids hand off a payload to one another on real hardware under the same object-centric abstraction.
Approach
decMHT introduces a common object-centric control abstraction that spans single-robot pickup, cooperative multi-robot transport, and robot-to-robot handover under one policy.
1. Object-Centric Attachment Specification
Each robot is given only a local attachment spec, rather than the payload's full geometry: an attachment-frame pose, lateral hand-contact offsets w, and a frontal clearance d. These three quantities define where a robot stands and where its hands make contact. The elongated box below makes the point visually: no matter how far the payload extends, each robot only ever needs its own local slice of it.
2. Hierarchical Policy Architecture
decMHT uses the same two-level hierarchy as our earlier quadruped work, decPLM. A trainable high-level loco-manipulation policy (2×128 MLP) maps the attachment spec, proprioceptive state, commanded attachment-frame velocity and height, and a binary mode flag to arm motor targets and a desired base velocity and height. A frozen low-level locomotion policy (2×128 LSTM) converts that into leg motor targets. Both run at 50 Hz. Every robot executes its own copy from purely local inputs, so no inter-robot communication is needed.
3. Mode-Gated Reward Structure
One shared policy expresses two behaviors, idle and track, through a mode-gated reward rather than separate networks.
Idle Mode
- Maintain zero commanded base velocity
- Maintain nominal torso height
- Ignore payload-motion tracking commands
- Return arms to nominal rest configuration
Track Mode
- Approach the assigned attachment frame
- Align both hands with their contact targets
- Hold and stabilize the payload
- Track payload velocity, yaw rate, and height
"No separate handover reward or handover-specific policy."
4. Handover Through Mode Switching
Handover needs no new reward or policy, only a sequence of idle and track modes applied to both robots, shown below. These mode transitions are also sampled during training, so the policy experiences the coupled dynamics of a live handoff and learns robust load transfer between robots.
5. Training
decMHT is trained with Independent PPO (IPPO): a shared policy is optimized on single-robot experience, then fine-tuned on two-robot experience, and deployed as independent copies using only local observations. Training follows a six-stage curriculum, from assisted contact through distant spawning to two-robot fine-tuning. Throughout, each robot's idle/track flag is switched randomly mid-episode, exposing the policy to every transition it needs at deployment, including joining, cooperating, and leaving through a handover, without any of these being trained as a separate, explicit behavior.
Key Contributions
- We introduce decMHT, a high-level, object-centric abstraction for multi-humanoid loco-manipulation that supports variable team sizes, attachment layouts, and object geometries.
- We develop a hierarchical decentralized control architecture, together with a training curriculum and reward design, that enables humanoids to coordinate through the shared object using only local observations and without direct communication.
- We demonstrate in simulation that a single trained controller supports teams of up to ten humanoids across diverse object geometries and attachment configurations, including robot-to-robot handovers across multiple elevation differences.
- We demonstrate sim-to-real transfer of the approach on one and two Digit V3 humanoids performing collaborative object manipulation.
For full quantitative results, ablations, and analysis, please refer to the paper. Simulation and hardware demonstrations of decMHT are shown in the videos above.
BibTeX
@misc{pandit2026decmht,
title = {Learning Multi-Humanoid Pickup and Transport via Decentralized Object-Centric Control},
author = {Pandit, Bikram and Gadde, Mohitvishnu S. and Shrestha, Aayam Kumar and Fern, Alan},
year = {2026},
eprint = {2609.17824},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2609.17824}
}