Nimi

English

中文

English

中文

Appearance

Generated Motion and Mapping Sidecars ​

Avatar generates backend presentation output from validated model capabilities and built-in routes. The current generated-motion catalog targets VRM. Live2D relies on its own built-in mappings and model capabilities; these VRM sidecars do not admit routes for Live2D.

Current VRM Path ​

Runtime supplies admitted semantics (such as agree), and Avatar maps them to catalog routes (such as nod_yes). The current routes are idle_subtle, listen_lean, nod_yes, shake_no, and greet_wave. These are Avatar backend IDs, not public APML tags or new Runtime activities.

The VRM provider takes the model, route, intensity, and loop inputs, returning either a generated clip or a fail_closed result with a reasonCode. When a capability profile is missing, a target is unsupported, or a pose is unsafe, the provider fails closed rather than silently falling back to idle. These routes do not require a physical .vrma file.

Exact Sidecar Fields ​

The current VRM parser accepts only these top-level fields; any unrecognized fields are rejected:

FieldPurpose
sidecar_idNon-empty mapping identifier
route_idAdmitted VRM route from the current catalog
backend_kindMust be vrm for this parser
profile_idMust match the capability profile used for support evaluation
confidenceValue from 0 to 1; support requires it to meet or exceed threshold
thresholdValue from 0.82 to 1; may raise, but not lower, the current confidence floor
manual_confirmationunconfirmed, confirmed, or rejected
source_kinddeterministic_manifest, static_name_match, llm_semantic_match, or human_review
target_fieldsNon-empty list; each target has target_kind and name, with an optional role

The schema contains no route or fallback fields. Confidence controls admission, not candidate ranking. Supply one unambiguous mapping per route rather than relying on ordering among duplicate route entries.

An LLM-derived mapping requires confirmed status. Rejected or below-threshold mappings remain unsupported. Valid values for target_kind are humanoid_bone, expression_preset, lookat, motion_group, and parameter_id; all declared targets must match the facts of the selected backend and capability profile.

Structural Example ​

This YAML can pass structural schema parsing. However, demo-vrm-profile and the neck target must still match an actual validated profile before the route is treated as supported:

yaml
sidecar_id: demo-nod
route_id: nod_yes
backend_kind: vrm
profile_id: demo-vrm-profile
confidence: 0.95
threshold: 0.9
manual_confirmation: confirmed
source_kind: human_review
target_fields:
  - target_kind: humanoid_bone
    name: neck

Passing structural parsing does not mean the model has executed the motion. Full support evaluation verifies the profile, backend kind, declared route, confirmation status, and target facts.

Source Basis ​

Documentation for Nimi — the installable, open-source, local-first personal AI product.

Copyright 2026-present Nimi Platform