Internal Animation Format

Overview

From a programming perspective, there are several steps to uploading an animation from a BVH (BioVision Hierarchy) file:

1. Read and parse the BVH file, creating an LLKeyframeMotion object containing the motion data.
2. Gather input from the user (via the upload preview floater) for things like animation priority, facial expression, and looping; these settings are stored in the LLKeyframeMotion object.
3. Serialize the LLKeyframeMotion object as LLSD.
4. Upload the serialized data to the asset server.

Before other viewers can play an animation, they must:

1. Download the serialized data from the asset server.
2. Deserialize it to an LLKeyframeMotion object.

Relevant source files:

• linden/indra/llcharacter/llbvhloader.cpp
• linden/indra/llcharacter/llkeyframemotion.cpp.

Note: binary file write is little Endian.

Header

The first part of the animation data is a header describing various details about the animation as a whole. The elements, in order, are:

field name	description	LLSD type	C++ type	Note
version		integer	U16
sub_version		integer	U16
base_priority		integer	S32
duration		real	F32
emote_name		string	std::string
loop_in_point		real	F32
loop_out_point		real	F32
loop		integer	S32	0: not looped, 1: looped
ease_in_duration		real	F32
ease_out_duration		real	F32
hand_pose		integer	U32
num_joints		integer	U32

Joint Data

After the header is data for each joint in the skeleton: Note: Unused bones need not be included in the file.

field name	description	LLSD type	C++ type	Note
joint_name		string	std::string
joint_priority		integer	S32

Joint Rotation Keys

At the start of the rotation data for each bone is the total number of rotation keys: If the bone has no rotation based keyframes, this value must be 0.

field name	description	LLSD type	C++ type	Note
num_rot_keys		integer	S32

Then, for each rotation key:

Note: These three values X Y Z appear to be the first three values of a truncated quaternion with the W term being calculated afterwards. Since a quaternion is X² + Y² + Z² + W² = 1 as long as you assume the W term has a consistent sign the X Y Z terms will be accurate.

field name	description	LLSD type	C++ type	Note
time		integer	U16	0: first frame, 65535: last frame
rot_angle_x		integer	U16	0:-1, 32767:0, 65535:+1
rot_angle_y		integer	U16
rot_angle_z		integer	U16

Joint Position Keys

At the start of the position data is the total number of position keys: If the bone has no position based animations, this value must be 0.

field name	description	LLSD type	C++ type	Note
num_pos_keys		integer	S32

Then, for each position key, position data is measured from Avatar Center ( mPelvis ), not joint resting position ( IE [0,0,0] is not joint resting position, it'll be the bone's position in parent coordinate space) with the exception of the mPelvis bone which is stored in world space coordinates.:

field name	description	LLSD type	C++ type	Note
time		integer	U16	0: first frame, 65535: last frame
pos_x		integer	U16	0: -5M, 32767: 0M, 65535: +5M
pos_y		integer	U16	pos measured from AV COG not joint offset
pos_z		integer	U16

Constraints

After the joint data are a number of entries for joint constraints. Constraints can target an avatar's parts in relation to each other or the ground.

field name	description	LLSD type	C++ type	Note
num_constraints		integer	S32

Then, for each joint constraint:

field name	description	LLSD type	C++ type	Note
chain_length		integer	U8	number of attached joints to include
constraint_type		integer	U8	0: point, 1: plane
source_volume		string	U8[16] array	skeleton collision volume name
source_offset		string (?)	LLVector3
target_volume		string	U8[16] array	skeleton collision volume name
target_offset		string (?)	LLVector3
target_dir		string (?)	LLVector3	value is currently ignored
ease_in_start		real	F32
ease_in_stop		real	F32
ease_out_start		real	F32
ease_out_stop		real	F32