Multiple failure sources
Failures can emerge from objective selection, robot–terrain interaction, strategy authority, or low-level execution.
Why decompose terrain adaptation?
The same locomotion task can fail for different reasons across terrain. TRACER asks a structural question: which component should adapt to a given terrain-induced failure?
Adaptation occurs at multiple levels
Existing approaches adapt different parts of the stack, but the roles and interfaces between them are often coupled.
Reward semantics are not physical semantics
Increasing a reward weight does not guarantee improvement in the physical quantity carrying the same label.
ηDesired physical preference
→
βInternal reward conditioning
→
ΘLocomotion strategy
→
JObserved physical outcome
TRACER makes these interfaces explicit so that a mismatch can be localized rather than hidden inside a single adaptation mechanism.
TRACER functional hierarchy
TRACER separates heterogeneous-terrain adaptation into four functional roles: Objective and Interaction provide upstream information, Strategy produces interpretable locomotion references, and Execution physically realizes the requested motion.
Core hierarchy used to diagnose where task preference, interaction information, high-level strategy, and low-level physical realization enter the locomotion stack.
ObjectiveWhat should be prioritized? Maps task, preference, and context into internal conditioning.
InteractionWhat dynamics are being experienced? Summarizes recent command–response history.
StrategyHow should the robot move? Produces interpretable high-level references such as velocity and body height.
ExecutionHow is the request physically realized? Tracks the strategy under dynamics and contact constraints.
Preference-conditioned physical outcomes
Changing task preference shifts the robot toward distinct physical operating points. The shifts are meaningful, but they are not perfectly axis-aligned.
Motion-biased
+1.45%
forward progress
20 / 20 paired initial conditions favorable
Stability-biased
−28.5%
slip
20 / 20 paired initial conditions favorable
Energy-biased
−2.99%
energy cost JE
19 / 20 paired initial conditions favorable
Controlled preference sweep on the same nominal plant with 20 paired initial conditions. Only task preference η changes.
Important: Energy-biased preference lowers JE relative to Balanced, but does not globally minimize it. This is one example of why reward semantics and physical semantics should not be treated as interchangeable.
What the initial study tells us
| Observation | Interpretation |
|---|---|
| Objective conditioning changes Strategy | Internal conditioning is behaviorally meaningful. |
| Physical effects remain coupled | Reward semantics ≠ physical semantics. |
| Stability is harder to isolate | The Strategy interface needs richer contact / gait authority. |
Current work: role-specialized adaptation
The interface diagnosis motivates a more explicit separation of continuous task semantics from interaction-driven structural adaptation.
Task / semantic path
η → Preference Calibration → β
η, β → Meta-Action Generator → Θᶜ
Task preference is retained as explicit semantic information while calibrated objective conditioning provides an internal coordinate.
Interaction / structural path
H → Response Adaptation → (ρ, σ)
(ρ, σ) → Gait Parameter Generator → D_req
Interaction history is routed to a structural gait channel, separating hidden physical mismatch from task intent.
Observed limitation
Physical preference ≠ reward semantics
Physical preference ≠ reward semantics
→
Extension
Preference calibration
Preference calibration
Observed limitation
Interaction not explicit
Interaction not explicit
→
Extension
Interaction adaptation
Interaction adaptation
Observed limitation
Limited Strategy authority
Limited Strategy authority
→
Extension
Structural gait authority
Structural gait authority
Items in this section extend beyond the initial workshop study and summarize the current TRACER direction.
Resources
Project materials for TRACER. This page will be updated with additional videos, code, and future publications.