Raibert Heuristic
- Also called: Raibert foot placement
- Category: Control › Locomotion control
- Tags: legged-locomotion, foot-placement, quadrupeds, balance-control, gait
A foot-placement rule for legged robots that puts the touchdown point ahead of the hip by half the stance duration times body velocity, plus a correction proportional to the velocity error.
Definition
The Raibert heuristic is a rule for choosing where a legged robot should place its foot so that it keeps moving at a desired speed without falling. It comes from Marc Raibert's work on hopping robots at Carnegie Mellon and MIT in the 1980s, described in his book Legged Robots That Balance (1986). Raibert observed that if the foot lands at the neutral point, the point the hip will pass over at the middle of stance, the body's forward speed is roughly unchanged by the stance phase, because the leg brakes the body in the first half of stance about as much as it pushes it in the second half. For a body moving at velocity v with stance duration T, the neutral point lies about v·T/2 ahead of the hip.
To change speed, the foot is placed away from the neutral point: further ahead to slow down, further behind to speed up. The form most used today is p_foot = p_hip + v·T_stance/2 + k(v − v_desired), where k is a feedback gain. Raibert's original controllers split hopping into three largely independent tasks: hopping height regulated by thrust during stance, forward speed regulated by foot placement during flight, and body attitude regulated by hip torque during stance. The foot-placement part carried over from one-legged hoppers to bipeds and quadrupeds, where legs that move together in a gait are treated as a single virtual leg.
The heuristic is still widely used in quadruped controllers. Model predictive controllers for robots such as the MIT Cheetah family compute ground reaction forces with MPC and place feet with the Raibert rule, sometimes with an extra term for turning. Learned locomotion systems use it as a reference for foot positions or as the low-level stepping rule beneath a learned policy that selects gaits. It assumes flat ground and a roughly symmetric stance and it ignores where safe footholds are, so on rough terrain it usually provides the nominal target that a foothold planning step then moves to the nearest safe location.
Why it matters
The heuristic turns a hard balance problem into a simple, tunable rule that works across many robots and gaits. It remains the default foot-placement baseline in quadruped controllers and a building block inside both model-based and learned locomotion systems.
Key applications
- Foot placement in quadruped trotting, bounding and pacing controllers
- Speed regulation for hopping and running robots
- Nominal footholds for terrain-aware foothold planners
- Reference foot positions for learned and residual locomotion policies
- Push recovery through corrective stepping
In practice
In a quadruped controller built around model predictive control, the gait scheduler decides when each leg swings. At each swing phase, the controller computes the touchdown position with the Raibert rule from the measured body velocity, the commanded velocity and the planned stance duration, then lifts the foot along a swing trajectory ending at that point, while MPC computes the ground reaction forces for the legs in stance. In hierarchical learning systems such as the one in the gait-transition work, a learned high-level policy chooses gait timing and the low-level controller places the feet this way. Engineers tune the velocity gain k: too small and the robot is slow to correct pushes, too large and stepping becomes oscillatory.
Common misconceptions
- The Raibert heuristic is a full locomotion controller — it only chooses foot placement; height, body attitude and contact forces need their own control
- It gives a good foothold on any terrain — it ignores terrain and assumes flat ground, so it is usually corrected by a foothold planner
- It is only of historical interest — it is still used in modern model-predictive and learned quadruped controllers
Who should know this
- Legged locomotion engineers
- Quadruped and humanoid controller developers
- Robot learning researchers working on locomotion
- Robotics students studying dynamic balance
Related technologies
- Capture Point
- Convex Model Predictive Control
- Gait Scheduler
- Swing Trajectory Generation
- Virtual Leg Model
Related terms
- Foothold Planning
- Spring-Loaded Inverted Pendulum
- Linear Inverted Pendulum
- Gait
- Quadruped
- Model Predictive Control
Related projects
- DimOS: Python robotics runtime with agents, SLAM and teleop, no ROS required
- PACE Sim2Real: Systematic sim-to-real transfer via actuator dynamics identification