Requested values
Relate the preference to terrain properties on a connected semantic-geometric graph.
The idea
Transfer the requested value.
Ground the route in the map and the robot.
Outdoor operators describe preferences such as avoid severe slopes or preserve energy. Yet an instruction alone does not determine a useful route: a suitable alternative may not exist near the start–goal query, and different robots experience the same terrain differently.
We present ongoing work on corridor-feasible language alignment. A semantic region graph exposes bounded alternatives, separating transferable mission values from environment- and embodiment-dependent route consequences. Preferences without meaningful local opportunity are reported as unsupported rather than inducing arbitrary detours.
Robot-specific Jackal and Spot cost models are trained using simulated traversal data. Initial held-out results support corridor-bounded adaptation from typed physical intents. Free-form language grounding and physical robot validation remain future work.
01 / Corridor contract
Bound the alternatives before selecting the route
A method-independent corridor is fixed using robot, distance, and objective-specific lower-bound routes. Candidate routes must satisfy the active robot’s mobility constraints and traversal-cost budget.
Feasible alternatives
Measure how much improvement is available relative to the neutral route.
Robot consequences
Evaluate complete routes using the active robot’s energy, risk and mobility model.
Insufficient opportunity
Make a lack of meaningful local alternatives explicit.
The prototype uses the CLEAR semantic-geometric abstraction to bind preferences and robot consequences to connected, elevation-aware terrain regions.
02 / Grounded routes
Same local opportunity. Different grounded response.
Terrain intent
“Prefer forested terrain while remaining robot-feasible.”
Robot-conditioned risk
“Minimize traversal risk for the active robot.”
03 / Initial evidence
Response, opportunity and robot feasibility
Held-out simulation evaluates typed slope, roughness and traction intents with separate wheeled and legged robot cost models.
Directional response
95% CI: 93.4–95.8%
2,519 actionable atomic cases
Mean CBIR
95% CI: 0.133–0.160
2,505 referenced cases
Within simulated budget
Every selected route satisfies the simulated robot-cost budget.
8 unseen terrain maps90 endpoint-pair clusters1,882 compound casesJackal + Spot embodiments
What counts as a response?
The preference gain must reach both an absolute floor of 0.005 and 10% of the available local opportunity.
Responseᵢ = 𝟙[gainᵢ ≥ max(0.005, 0.10 × opportunityᵢ)]
How much improvement is recovered?
Corridor-Bounded Intent Regret (CBIR) compares the chosen route with an approximate objective-specific reference from the same feasible alternatives.
CBIR = (g(chosen) − g(reference)) / (g(neutral) − g(reference))
0 = reference · 1 = neutral · lower is better. CBIR applies when the reference improves on neutral routing.
04 / Validation agenda
The next tests
Generalization
Independent maps and sensor conditions, including cases with no meaningful alternative.
Embodiment
Paired Jackal–Spot trials on the same maps and instructions, measuring physical outcomes.
Alignment
Blinded human route comparisons, including disagreement and inter-rater reliability.
Matched controls
Comparisons with adaptive search and existing language-conditioned cost-map methods.