TopoFocus Real-World

Real-world dual-robot implementation of TopoFocus. An RTX 4090 Hub builds the shared semantic map and coordinates VLM-selected goals; both robots plan, control and stop locally.

Real-world dual-robot platforms: Robot 0 (Unitree Go2) with RealSense D435i, Robot 1 wheeled chassis with an Odin1 spatial memory module

TopoFocus system architecture: shared semantic perception and VLM coordination on the RTX 4090 Hub with robot-local planning, control and safety

The Hub fuses online maps and applies the source-compatible A–D VLM decision cascade. Each robot routes an expiring goal with known-free A*, then follows a TinyNav local trajectory behind independent fail-closed safety checks.

Use the same Git revision on every host.

git clone https://github.com/AlanZhu2006/topofocus_realworld.git
cd topofocus_realworld
python3 hub/tools/verify_public_baseline.py --workspace .

Checks the public file and deployment contracts only; it installs nothing and does not connect to a robot.

bash hub/scripts/bootstrap_gpu_hub_cleanroom.sh
bash hub/scripts/bootstrap_gpu_hub_cleanroom.sh \
 --apply --fetch-models --accept-model-licenses

The first command previews the plan. The second installs the locked runtime, pinned model artifacts and validation stack.

bash hub/robot_overlay/bootstrap_robot0_cleanroom.sh
bash hub/robot_overlay/bootstrap_robot0_cleanroom.sh --apply
bash hub/robot_overlay/configure_go2_network.sh
bash hub/robot_overlay/configure_go2_network.sh --apply

Commands without --apply preview their plans. The applied commands install the locked Jetson runtime and dedicated Go2 network configuration without starting motion.

Full reproduction guide · Robot 0 clean-room deployment · Supervised workflow · Documentation index

Shared cross-robot calibration using a circle-grid board

Both stationary robots observe the same ×ばつ10 circle-grid board to register their local odometry in one shared frame. A moved-board holdout independently validates the fitted transform before use.

Note: Glass, mirrors, low-texture surfaces and extreme lighting can degrade depth sensing and localization; adjust the viewing direction or use another scene, then revalidate tracking before a run.

The short-range setting starts both robots from the same lab area and targets a nearby white chair. All five formal experiments succeeded.

Standard SPL uses the independently measured shortest feasible path L≈3.25 m; source-compatible SPL uses the arriving robot's start-to-arrival displacement D as a source-compatible reference.

Experiment 1 explored semantic map and robot trajectories

Full five-experiment archive · Machine-readable results · Media manifest · Run instructions

The medium-range setting requires both robots to explore cooperatively toward a more distant plant target.

Both means count the two failures at zero contribution. Standard SPL uses the operator-provided independently measured shortest feasible path L≈7 m (Scene 01's 3.25 m is not reused).

Experiment 2 explored semantic map and robot trajectories

Formal 01 failure record · Formal 02 failure record · Formal 03 success record · Formal 04 record · Formal 05 record · Machine-readable results

The long-range setting retains the plant target while increasing travel distance and requiring both robots to explore and coordinate across the route.

Both means count the two time-limit failures at zero contribution. Standard SPL uses the independently measured approximate shortest feasible path L≈14 m.

Experiment 3 explored semantic map with Robot 0 and Robot 1 trajectories

Full five-experiment archive · Formal 01 failure record · Formal 02 failure record · Formal 03 success record · Formal 04 success record · Formal 05 success record · Machine-readable results · Media manifest

Five formal experiments are archived. Robot 1 reached the plant in Formal 04 and Formal 05, with Robot 0 contributing its independently allocated exploration routes.

The source-compatible mean counts all three failures at zero contribution. Standard SPL uses the operator-provided independently measured approximate shortest feasible path L≈11 m; its mean also counts the three failures at zero contribution.

Experiment 4 explored semantic map and robot trajectories

Formal 01 failure record · Formal 02 failure record · Formal 03 failure record · Formal 04 success record · Formal 05 success record · Machine-readable results