Embodied AI Field Map

中文版

Embodied AI is a closed-loop systems field, not a single model family. This map separates capability coverage from repository evidence so that a topic name never implies a reproduced result.

Capability stack

Capability	Engineering question	Best foundation	Pipeline contract
Sensing & calibration	Are observations synchronized, calibrated, and healthy?	Perception and sensors	Perception & state estimation
State estimation	What state is safe to use, with what uncertainty?	Probability and optimization	Perception & state estimation
Data & simulation	How are tasks, demonstrations, splits, and perturbations produced?	MuJoCo · Dataset and training	Simulation & data
Policy learning	How does observation and language become action?	Deep learning · Transformers	VLA policy
Predictive models	What will happen after an action?	Probability and optimization	World model planning
Interactive improvement	How should behavior improve from reward and failure?	Control	RL post-training
Generalist policies	How are models adapted across datasets and bodies?	Dataset and training	RFM & cross-embodiment
Task reasoning	How is a long instruction decomposed and replanned?	Robot systems and safety	Embodied reasoning
Manipulation & dexterity	How are geometric or learned commands mapped to constrained motion?	FK, Jacobian, and IK	Dexterous retargeting
Navigation & locomotion	How does an embodiment move while remaining localized and stable?	Control · Systems and safety	Navigation & locomotion
Transfer & deployment	What must pass before risk is increased?	Evaluation and reproducibility	Sim-to-Real

Evidence today

Level	Tracks	Meaning
Smoke-tested	Simulation/data, VLA, world model, RL, dexterous retargeting	A lightweight repository path completes; performance is a separate claim.
Interface-tested	RFM/cross-embodiment, embodied reasoning	Local schemas, adapters, or planners connect without proving real weights or hardware.
Documented	Sim-to-Real, perception/state estimation, navigation/locomotion	The engineering contract and gates exist; no universal local command represents the system.

Choose by research goal

Goal	Start	Then prove
Learn robot learning from zero	Foundations overview	Complete one smoke-tested pipeline and retain its artifacts.
Build a multimodal policy	VLA pipeline	Closed-loop success, language ablation, latency, and failure cases.
Study prediction and planning	World-model pipeline	Multi-step rollout error and planned task success separately.
Work across robot bodies	RFM pipeline	Action semantics, adapter coverage, and per-embodiment results.
Study dexterous hands	Retargeting pipeline	Geometry, temporal quality, contact/task evidence, and hardware evidence separately.
Build mobile or legged systems	Navigation/locomotion contract	Localization, tracking, collision/fall, recovery, and transfer evidence.

Deliberate non-claims

The repository does not currently claim a reproduced SLAM benchmark, navigation success rate, legged-locomotion policy, general-purpose hardware deployment, or competitive large-scale foundation-model result. These are visible expansion targets, not hidden assumptions.