Local and remote deployment
Both modes use the same actor/learner protocol and durability; remote mode adds an encrypted tunnel because
the bearer token does not encrypt traffic.
Local train · one Windows host
Remote actor · separate Trackmania host
Shared learner runtime
spawn
rollouts
CLI launcher
multiprocessing
spawn
Actor process
game · policy ·
spool
Learner process
loopback gRPC
Actor + spool
stable actor ID
SSH or VPN
authenticated
encryption
Loopback
port
SSH/VPN
forward
Same run identity
token · fingerprint ·
map · contracts
Coordinator
endpoint
bounded authenticated
requests
WAL + applied
frontier
persist before ACK
Replay + learner
sample · update ·
evaluate
Portable outputs
policy snapshot ·
checkpoint
Local does not mean in-process: actor and learner are separate spawn-safe processes, and failures must
propagate to the launcher.
Never expose gRPC directly. Keep the learner on literal loopback and carry it through an authenticated
encrypted tunnel.
Download editable diagram