Motivation

The trick behind a moving character

• a surface that looks like a person or creature
• an internal skeleton that gives it structure
• motion data that says how the body changes over time
• rendering choices that make the result readable

Four-Week Timeline

Week	Mode	Focus
Week 1	Individual	Implement and debug forward kinematics on a simple block character.
Week 2	Individual	Implement and compare one-hot skinning and linear blend skinning on SMPL.
Week 3	Pairs	Explore human motion clips and reconstruct group-member characters.
Week 4	Pairs	Refine the group animation scene and export the final video.

Assessment

The Project In One Sentence

Build a small character-animation pipeline, then use it to create an animated human-avatar scene.

The Pipeline

Build a small character-animation pipeline, then use it to create an animated human-avatar scene.

This project studies a simple version of the character-animation pipeline:

1. represent a character as a skeleton and a mesh
2. pose the skeleton with forward kinematics
3. attach a surface to that skeleton with skinning weights
4. turn a sequence of poses into a short animation

The emphasis is on understanding the mechanics of animation, not on learning a large authoring package.

Materials

• docs/: handouts, report requirements, and setup instructions
• viewer/: interactive viewer and scene tools
• viewer/student_submission/: files students edit
• assets/: block characters, motions, and SMPL assets
• slides/: browser presentation decks and local notes support

Week 1: Forward Kinematics

From skeleton hierarchy to motion

Why We Start With Blocks

The early parts use a blocky articulated character rather than a realistic human mesh.

That is deliberate. A block character makes the hierarchy easy to see:

• each body part is rigid
• each joint is easy to identify
• FK bugs are visually obvious

Once the skeleton logic is clear, we switch to SMPL and study mesh deformation.

Part 1: Forward Kinematics

• explain what a joint hierarchy or kinematic chain is
• distinguish local coordinates from world coordinates
• compute world-space joint rotations and positions from local transforms
• debug a broken articulated character by checking parent-child relationships
• create a short custom motion by placing key poses on a timeline

Output: one custom motion clip with at least 10 keyframes.

Demo: Show The Viewer

• skeleton overlay
• joint selection
• shoulder or hip rotation
• timeline keyframes
• exported motion video

Viewer Controls

• Motion: choose a built-in or custom motion clip.
• Animate: play or pause the selected motion.
• Show Skeleton, Show Mesh, Show Skinning Weights: toggle visual layers.
• Selected Joint and Joint Editor: inspect and edit one joint.
• Timeline: capture keyframes for a custom sequence.
• Export Motion Video: render evidence from the current browser camera view.

Week 2: Skinning And SMPL

From skeleton motion to deforming surfaces

FAQ

Open the FAQ

• setup and Git questions we expect more than once
• short answers collected in one student-facing page
• updated as recurring questions come up

Current FAQ Items

• conda can be used in place of mamba
• use two Git remotes if you want to keep custom code in your own repository
• pull course updates from the teaching repo; push your work to your repo
• SMPL model download details are in the Part 2 notes
• Windows SMPL loading issue: see FAQ and GitHub issue #1

Why Skinning Changes The Problem

• same FK poses from Part 1
• a human mesh that bends around joints
• weights decide how much each bone moves each vertex
• one-hot binding is the useful baseline; LBS is the smoother version

Part 2: Skinning And SMPL

• explain what a skinning weight means
• derive one-hot skinning from a full weight matrix
• explain why one-hot attachment creates artifacts near elbows, shoulders, hips, and knees
• compare one-hot binding against linear blend skinning (LBS)
• reuse the same motion on both the toy character and SMPL

Same motion, different character surface.

Inputs To Skinning

• model_data.rest_vertices: mesh vertices in the default pose, shape (V, 3)
• model_data.rest_joints: joint centres in the default pose, shape (J, 3)
• model_data.ground_translation: shift that puts the rest body on the floor
• world_rotations: posed joint rotations from FK, shape (J, 3, 3)
• world_positions: posed joint centres from FK, shape (J, 3)
• model_data.skinning_weights: vertex-to-joint weights, shape (V, J)

Local Coordinates

Local coordinates are measured in a moving parent frame.

• joint.translation: offset from parent joint to child joint
• local_rotations[j]: how joint j rotates relative to its parent
• if the parent turns, the child's local axes turn with it

World Coordinates

World coordinates are measured in the fixed viewer scene.

• one shared origin and axis system for the whole character
• world_positions[j]: final scene position of joint j
• world_rotations[j]: final scene orientation of joint j
• the renderer draws points after they are in world coordinates

FK Converts Local To World

For a child joint j with parent p:

world_rotations[j] = world_rotations[p] @ local_rotations[j]
world_positions[j] = world_positions[p]
                     + world_rotations[p] @ joint.translation

Why Skinning Uses World Transforms

Local rotations are the controls; world transforms are the result.

• a hand vertex is affected by shoulder, elbow, and wrist motion
• the wrist's local rotation alone does not include its parent joints
• after FK, each joint has a final world position and orientation
• skinning uses that final joint frame to move nearby vertices

Rest Vertices And Rest Joints

The rest pose is the character before animation.

• rest_vertices[i]: where mesh vertex i starts in rest pose
• rest_joints[j]: where joint j starts in rest pose
• add ground_translation so both are in viewer world coordinates
• rest_vertices[i] - rest_joints[j]: vertex i as seen from joint j

World Positions And Rotations

Forward kinematics has already accumulated the hierarchy.

• world_positions[j]: where joint j is in the current pose
• world_rotations[j]: how joint j is oriented in the current pose
• these are the posed world-space transforms from Part 1 FK

The Skinning Transform

Each joint gives vertex i one possible posed location.

• start from the vertex's rest-pose offset from that joint
• rotate that offset using the joint's posed world rotation
• place the rotated offset at the joint's posed world position
• combine the joint proposals using the skinning weights

One-Hot Skinning

One-hot skinning is the simplest baseline:

• find the joint with the largest weight for each vertex
• set that joint's weight to 1
• set all other weights to 0

This creates rigid piecewise motion. It is easy to implement and easy to interpret, which is why it is a good first baseline.

Linear Blend Skinning

In LBS, a posed vertex is a weighted sum of the transforms from several joints:

v'_i = sum_j w_ij T_j v_i

You do not need to derive the full production form of LBS in this part, but you do need to understand the idea: vertices near joints are shared across influences instead of being assigned to a single bone.

Reports

• Interim report and animation results: Friday 29 May 2026, 2pm.
• Part 3 showcase: Tuesday 9 June 2026, 11am-1pm, LT6.
• Final report and animation results: Friday 12 June 2026, 4pm.

• your code
• your saved custom motion clip
• a short video of the toy-rig motion
• a short video of the same motion on SMPL
• one-hot versus LBS comparison figures
• the interim report PDF

Submit on Moodle

Part 3

Work in groups of two to create a coherent 30-second (or longer) animated human-avatar scene.

• explore human motion sequences using a human body model
• understand how SMPL-compatible motions can drive a skinned character
• connect a reconstructed character mesh to the same SMPL-24 body skeleton
• choose, combine, and refine motion clips for a short scene
• produce a final 30-second (or longer) animation video
• explain the strengths and limitations of your character-animation pipeline

Preview: Scene Composition

• preset human motion clips
• multiple character tracks
• root placement and facing
• draft preview before final export
• saved .scene.json files

How Students Should Work

• implement a small piece
• test it visually in the viewer
• save evidence immediately
• compare failure cases, not just final results
• ask for help with a specific screenshot or clip
• use the Git starter if Git commands are new

AI Use Policy

• Parts 1 and 2: no AI-generated code, math derivations, or results.
• Part 3: AI tools may support creative/production work.
• Final focus: a realistic, controllable 3D avatar rendering scene.
• Reports: no wholesale AI-written content; grammar help is allowed.
• Every report needs an AI Use Statement.

Where To Go Next

• Setup
• Part 1
• Part 2
• Interim report
• Part 3
• Final report
• FAQ
• References
• Project repository