G1 corridor walking: does a 5 cm entry shift change wall contact?

After making a video of G1 clearing an 80 cm corridor, we changed the question: can it still pass when it starts slightly off-centre? For a deployment study, knowing the conditions that fail is more useful than one successful example.

The purpose is to test whether corridor width, entry position and heading change wall contact under a fixed walking controller, and to measure failures that reaching the exit alone would miss. This does not determine a minimum corridor width for hardware.

The conclusion is that this controller can make wall contact depending on entry conditions, despite reaching the exit. These results do not establish readiness for deployment. Contact-producing entry conditions are the next target for controller improvement.

We ran 36 Isaac Sim trials with different corridor widths and entry positions. At the same 80 cm width and entry angle, the central start passed, while a 5 cm lateral shift produced wall contact. Width alone did not determine the operating range of this controller.

All 36/36 trials achieved full-body exit. Once contact and conservative clearance were included, only 18/36 met the complete criterion.

The same width, a different entry position

Both videos use a forward command of 0.5 m/s, initial heading -2°, and seed 100. Only the starting lateral position changes. Each recording plays at real simulation speed; the contact case is also shown through its end.

In the corresponding matrix trials, the central start measured 0 N wall force and a conservative full-body clearance lower bound of 29.0 mm. Starting at -5 cm produced a maximum resultant wall force of 691.9 N. Both robots fully cleared the exit. A reach-the-goal criterion would hide this difference.

This does not mean every run with these starting positions will behave this way. The complete set of conditions is shown below.

Width and entry-position results

We tested internal widths of 70, 80, 90 and 100 cm, with starting lateral positions of -5, 0 and +5 cm relative to the corridor centre. Each cell contains three initial headings: -2°, 0° and +2°, paired with seeds 100, 101 and 102. The same combinations were reused at every width and position.

The table counts trials that met our criterion. Three trials per cell means three small heading perturbations, not three identical resets. These counts are not estimates of real-world success probability.

Scroll horizontally to see the full table.

None of the four tested widths passed all nine trials. Untested intermediate widths, other floor conditions, payloads and speeds remain outside that result.

We also separated the reasons for not passing:

Scroll horizontally to see the full table.

At 1 m width, all nine trials recorded 0 N wall force. One trial remained uncertified because its enclosing-box clearance lower bound was approximately -5.5 mm. This is not evidence of wall contact at 1 m: it identifies a need for finer geometric evaluation or a larger control margin before accepting the result.

“Wall contact” means force was detected by a wall sensor. “Clearance not certified” means no wall force was detected, but the conservative boxes enclosing the body did not maintain a positive margin. Overlapping boxes alone do not establish physical contact. A trial with both contact and an insufficient bound is classified as contact.

What counts as clearing the passage?

We track 51 bodies with geometry, including shoulders, arms, fingers and legs. All enclosing bounds must extend beyond the wall end by at least 5 cm, followed by at least 0.5 seconds of continued walking. We also require no detected wall force, positive conservative full-body clearance and no fall. Stopping at the exit is not included.

Both walls record resultant force at 200 Hz. At each step, saved body poses let us calculate the clearance of enclosing geometry boxes. This is a lower bound, not an exact surface-to-surface distance. We recomputed clearance and exit from saved poses and matched the online records for all 36 trials.

The wall sensors were checked using an earlier contact-positive control. Detailed contact-point callbacks were not validated in this runtime, so their empty logs are not used as evidence of no contact.

Implementation

We use a publicly released Unitree walking policy and its matched model, with no new policy training. Our controller calculates the desired path and lateral correction in world coordinates, then converts them to body-frame velocity commands. It does not reposition the robot root or animate its feet.

Physics runs at 200 Hz and the policy at approximately 33.3 Hz. The first two seconds have zero forward command, followed by a one-second ramp to 0.5 m/s. The walls are 4 m long and 1.5 m tall; trials last at most 20 seconds. Gains and model settings were not tuned between matrix trials.

Position and heading come from simulator state. This is not a Visual SLAM, Nav2 or hardware locomotion result.

How to use this in a deployment study

The table describes the tested model and controller. It is not a specification for the corridor width required by a physical robot.

When reproducing a site's passage, include entry position and angle as well as width. In our 80 cm comparison, a 5 cm change in starting position changed whether contact occurred. This does not establish a 5 cm localization-error tolerance: starting offset and ongoing localization error require different tests.

The next questions concern carrying posture, floor conditions, stopping and turning, and using visually estimated poses. Keeping the failed conditions helps decide whether to constrain entry, improve control, or test a wider passage.

For a similar evaluation, check wall contact and whole-body clearance as well as arrival, and vary entry conditions. The conditions and criteria in this article provide a starting point for that evaluation. The walking policy comes from the Unitree upstream implementation.

Team Z helps development teams connect site layouts and motion requirements to simulation, implement evaluation code and investigate failed conditions. Discuss your development needs if you have a passage or motion to evaluate.