search

google-deepmind / mujoco_mpc

Real-time behaviour synthesis with MuJoCo, using Predictive Control
https://github.com/deepmind/mujoco_mpc
Apache License 2.0
952 stars 142 forks source link

Multi-agent MPC #334

Closed hosseinfeiz closed 1 month ago

hosseinfeiz commented 1 month ago

Hi,

In the case of multi-humanoid tracking, is it possible to instantiate several agents, associate separate planners to each agent, and use ILQG to minimize the tracking objectives while considering future interaction contacts? I assume a higher-level agent is needed to coordinate the contacts between multiple agents.

thowell commented 1 month ago

To consider interaction between humanoids, one option would be to create a single MuJoCo model that contains $N$ humanoids and use the existing iLQG planner to jointly optimize all humanoid trajectories. With this setup, interactions between humanoids will be considered. However, the planning problem will have complexity $O(N^3 n_u^3 T)$.

Alternatively, as you suggest, individual planners could be used for each humanoid, but then interaction between humanoids would need to be handled by a separate mechanism. One option would be to consider the motions of the other humanoids fixed (they could be modeled as time-varying but fixed geoms in the scene), and optimize trajectories for a single humanoid at a time. Iterating this scheme over all humanoids might converge to a reasonable result.

Both of these options could use MJPC as is, without a new/custom planner.

hosseinfeiz commented 1 month ago

Thanks, it worked perfectly.