MuJoCo obstacle clearance: a point path is not a body path

git clone https://github.com/VirtualSelect/hohoo-embodied-agent.git hohoo-embodied-agent-study
cd hohoo-embodied-agent-study
git checkout 831789c2a1efb4564c4cfabc2298e48e16ada879
python -m venv .venv
$py = '.\.venv\Scripts\python.exe'
& $py -m pip install -r requirements-lock.txt
& $py experiments/planar_reach/run.py --out outputs/my-run
& $py experiments/planar_reach/audit.py outputs/my-run

Pinned code · Raw evidence · Lab record

The empty scene reaches its target. Add a box at x=0.3m while keeping the target at x=0.6m and the controller unchanged: mass 1kg, Kp=80, Kd=18, ±5N per axis. A direct path hits the wall. A path that merely clears it for a point also fails because the actuator has a 35mm radius.

Fixed path comparison

The wall has x half-width 40mm and y half-width h of either 40mm or 90mm.

Route	Targets
direct	(0.6,0)
corner	(0.16,h), (0.44,h), then goal
clearance	(0.16,h+0.1), (0.44,h+0.1), then goal

The controller advances within 15mm of a waypoint. The 100mm clearance is a fixed teaching choice, not an optimized or certified safety distance.

For a sphere, expanding obstacles by its radius converts body clearance into a point-path question. Tracking error adds another margin: inertia can cut corners or overshoot. This example uses fixed waypoints, not a general planner.

Actual contact results

Wall half-height	Route	Contact samples	Final error	Completion
40mm	direct	1850	374.79mm	No
40mm	corner	1680	376.86mm	No
40mm	clearance	0	<0.01mm	Yes
90mm	direct	1850	374.79mm	No
90mm	corner	1670	385.19mm	No
90mm	clearance	0	<0.01mm	Yes

Contact samples are 2ms states containing contact, not separate collision events. Sustained contact contributes many samples. No impact-force or hardware-damage conclusion follows.

Both clearance routes pass; the other four fail. This supports the need for body clearance in this scene, not universal obstacle avoidance.

Inspect the right trajectory

Waypoints are controller targets; mj_step integrates actual position. The plot uses recorded x/y, not a drawing of intended targets. Direct motion should stop at the wall, corner motion should remain blocked, and clearance should travel above it. If a run passes through the wall, inspect contact masks before celebrating the planner.

Exercise: enlarge the sphere without changing waypoints. Compute the geometric clearance, account for the waypoint tolerance, then inspect actual contacts. Lower damping as a second exercise to see whether a geometrically valid path remains trackable.

Dynamic obstacles, articulated robot configurations, narrow passages and hardware safety are outside this study. Next: delayed noisy feedback.

Reproduce from a clean checkout

Python 3.12. The runner executes all three comparisons; each article discusses its own subset.

git clone https://github.com/VirtualSelect/hohoo-embodied-agent.git hohoo-embodied-agent-study
cd hohoo-embodied-agent-study
git checkout 831789c2a1efb4564c4cfabc2298e48e16ada879
python -m pip install -r requirements-lock.txt
python experiments/planar_reach/run.py --out outputs/my-run
python experiments/planar_reach/audit.py outputs/my-run
python experiments/planar_reach/plot.py outputs/my-run

Use a new output directory. Windows was exercised; Linux/macOS were not rerun. No model API or key is required. The source commit in the manifest precedes the evidence commit linked above; source hashes bind the executed files.