MODULE 40 / CHARACTER & EXPRESSION
Give the walk a character
Understand how stylized locomotion differs from playing a head clip. Build a repeatable simulation study, then identify the precise controller changes needed for an expressive gait.
By the end, you can…
- Compare the Disney RL approach with model-based gait authoring.
- Annotate support phases and evaluate one style change at a time.
- Know when a new command interface and retraining are required.
Modules 13–16 and 31: a reproduced baseline policy simulation and training workflow.
Modules 36–39. Hardware is not needed for this module.
DIRECT THE PERFORMANCE
Exports four semantic channels at 60 samples/sec in degrees. Desktop files need your calibrated joint map before simulation. Hold and linear modes demonstrate discontinuities.Two papers, two control strategies
The BD-X paper uses separately trained standing, periodic walking and episodic-motion policies. While walking, lower-body motion comes from the periodic policy; animation does not directly supply its leg trajectory. The animation system provides selected commands such as head/torso targets and velocities.
The stylized-gait authoring paper instead combines animation curves and contact timing with model-based control. Blending is aligned to gait phases, and the controller adjusts motion for balance. These are related creative goals with distinct technical implementations.
A foot cannot be everywhere at once
A support phase says which foot is carrying contact and which is moving. If two clips put the left foot in different phases, averaging their angles can drag a planted foot or put both feet in the air. Align the events before thinking about blending.
For a first study, annotate left support, double support, right support and double support when present in your chosen gait. Some gaits have different contact sequences; use observed contacts rather than forcing every gait into four equal quarters. Foot motion seen in video is evidence to inspect, not a direct contact-force measurement.
A trained policy only understands its contract
Inspect the baseline policy’s observation definitions and metadata. Does it accept head targets, torso orientation or a style parameter, or only a velocity command? An unused new number does nothing. Changing the input width without retraining breaks compatibility. Overwriting leg actions after inference bypasses the behavior the policy learned.
Head motion can also disturb the body. If the controller was trained with a fixed head, a large moving head may be out of distribution. Add a bounded reference to the simulation task, train or adapt a compatible tracker, then export matching metadata. Keep the original baseline and rollback intact.
Turn “looks better” into an experiment
Use the same forward command, duration, terrain and random seeds for the baseline and variant. Evaluate falls, foot slip, tracking error, joint saturation and timing alongside viewer judgments. A prettier motion that falls more often has not solved the task.
A first advanced training change might introduce one small head-yaw reference while preserving the rest of the task. The task must expose the reference to the policy, compute its tracking error, limit its sampled range and log the result. Confirm a short smoke run and the observation/action shape before budgeting a longer run. The training modules explain that environment-specific implementation; this kit supplies no drop-in Disney policy.
Where the newer research fits
Walk Like This separates a style reference from commanded velocity, generates a kinematic reference and tracks it with an RL controller. It is an advanced reading destination after you can run controlled comparisons. Do not make a diffusion model your first prerequisite for a small original gesture.
For this course, a successful first result is modest: a stable baseline and one measured style experiment. Larger changes in stride, torso sway or an added tail need a richer training and validation project.
WORDS YOU NOW KNOW
- Gait phase
- A location in a repeating locomotion cycle, usually tied to contact events.
- Tracking controller
- Feedback control that attempts to follow a specified reference.
- Out of distribution
- A situation unlike the ones a learned system was trained to handle.
YOUR CALL
Finished with this module?
0 of 7 lab steps checked. 0 of 2 questions answered. You decide when the module is complete. Revisiting or editing it will not change your decision.
You can reopen it later. Checking every box will not automatically complete it.Go to the source
Primary references for this module. Use the version and hardware revision named in the lesson.
Disney · Design and Control of a Bipedal Robotic Character · RSS 2024Read sections V–VI for policy commands and the animation engine. This is Disney BD-X, not the independent BDX-R hardware.Disney · Interactive Design of Stylized Walking Gaits for Robotic Characters · 2024Contact-aligned animation and model-based control; a different control method from the BD-X RL paper.Disney · Walk Like This · CoRL 2026Advanced reading: style reference + velocity command → generated reference → RL tracker. Not required for your first gesture.BDX-R · Pinned deployment sourceInspect this exact revision and the policy metadata used in your earlier labs before designing an adapter.