跳转至

Capability Matrix

最后更新:2026-07-27。此页只索引外部方法的论文原生能力、官方 release scope 和 PhysHSI 中的比较角色;它不等于本项目已经复现。全部来源和纳入规则见 Deep-Research Audit

K 表示 kinematic 或 soft-physics trajectory method;P 表示论文报告 physical humanoid closed loop。full/partial/absent 是截至审计日的官方 release scope,不是 论文质量判断。

方法 原生输入 → 输出 K/P hand/contact 语义 teacher → student 官方 release 已核验 PhysHSI 中的正确角色
HUMANISE / SceneDiffuser language + 3D scene → human motion/pose K scene contact,非 HOI tracker code/data static HSI generation lineage
TRUMANS / LINGO scene/action/text → long HSI;dynamic object motion K LINGO: SMPL-X/MANO code/data reference-source / generator comparison
InterDiff / CG-HOI human/object history or text → dynamic HOI K contact-informed code physical-informed generation baseline
CHOIS language + initial state + sparse object waypoints → synchronized human/object trajectory K object path/contact metrics code 最直接的 path-conditioned proposal baseline
HSI-GPT / SceneMI multimodal tokens or noisy keyframes + scene → HSI motion K SceneMI: BPS/local geometry paper/project scene interface / noisy-repair baseline
HOSIG / Dyn-HSI scene/text → full-body HOSIG / dynamic HSI K HOSIG finger-level; Dyn-HSI SMPL-X hands HOSIG promised; Dyn-HSI absent dynamic-scene generation comparison
MP-HOI multimodal priors/text → cascaded SMPL-X human/object HOI trajectory K contact/dynamics features; processed data omits articulated hands paper only / release unverified text/multimodal HOI generation comparison
SSOMotion / DIP scene semantics or unpaired scene → human motion K scene relation,非 HOI control code scene representation/generalization baseline
InterPhys / PAMotion dynamic scene/text → force/contact-aware HOI trajectory K continuous-force / multi-object contact loss not verified kinematic physics-aware comparison,不是 teacher
HOI-FHLI instruction → body/finger/object motion → RL tracker P full finger/object motion tracker only full end-to-end generator + direct tracker baseline
SIMS RAG script + scene → physics HSI policy P multi-condition physical contact policy only partial/placeholder instruction-to-physics system competitor
BiBo VLM command + executed motion → diffusion future → RL tracker P IK key-joints, not articulated finger replay executor only partial physical-feedback system; hand-interface boundary
VLM-RMD VLM relation plan → auto goal/reward → RL policy P part relation/movement policy only dataset/project only reward/goal-interface competitor
DeVI text-conditioned synthetic HOI video → hybrid target → dexterous physics controller P 3D human + 2D object hybrid tracking reward controller only official repo is code stub; release pending generated-video-to-physics HOI system competitor
PULSE / PULSE-X physics motion latent encoder/decoder → tracking/task action P full-body; original PULSE is 32D, project PULSE-X checkpoint uses a separately verified 48D contract no full source/preview frozen latent controller for A2-G, not a new method contribution
PHC-X checkpoint family dense SMPL-X reference → 153D direct action target P articulated SMPL-X body/hands no local released checkpoint contract direct lower bound and A2-R base; not a native ResMimic interface
MaskedMimic / ProtoMotions masked partial targets/text/scene → physics motion inpainting P released checkpoint is SMPL, no articulated hands no full source; SMPL checkpoint IsaacLab-only useful partial-reference prior, not an off-the-shelf SMPL-X hand HOI teacher
SONIC G1 dense motion reference → universal tracker P G1-specific 64D/FSQ + hand primitive in GRAIL no official code/models GRAIL native base, not contract-equivalent to PULSE-X
H2O / OmniH2O retargeted human motion → privileged teacher → sparse-sensor humanoid policy P dexterous hands, not dense HOI object trajectory yes full official code/data core human-to-humanoid teacher/deployment prior
HumanPlus human-motion shadowing → teleop data → egocentric BC skill policy P robot body/hands, task-specific objects yes paper/project; runnable release unverified full-stack human-data-to-visual-skill competitor
TWIST retargeted human/MoCap reference → RL+BC whole-body tracker P robot whole-body, not a dense object-contact teacher no full code/data/checkpoint open human-to-humanoid tracking/teleoperation baseline
Expert-Guided Imitation MoCap upper/root + walking-expert BC → box loco-manipulation P box/contact signals; no articulated SMPL-X hand interface expert-guided imitation paper/project; runnable release unverified direct human-reference + locomotion-expert comparison
PhysHOI / SkillMimic dense reference → scratch contact-aware policy P explicit interaction reward teacher only source lineage scratch expert lower bound
OmniGrasp object goal → grasp/control policy P object-centric policy only source lineage object-control diagnostic
InterMimic dense human/object/contact reference → HOI tracker P whole-body + object/contact teacher official source pretrained HOI teacher family
MaskedManipulator / MimicManipulator dense SMPL-X HOI reference → physics teacher → DAgger masked body/object-goal policy P articulated body/hands + phased contact reward yes paper/project; runnable release unverified direct teacher + sparse-goal student competitor; cannot be omitted from SMPL-X physics discussion
ResMimic frozen GMT + object/ref → direct action residual P object/contact reward teacher full A2-R inspiration; adapted, not strict PHC-X reproduction
GRAIL 3D asset/VFM → SMPL-X/object → SONIC latent adapter → RGB policy P 64D latent residual + binary hand primitive yes, visual policy full code/data A2-G design source and direct system comparator
RePHO noisy HOI → forward/backward policy/reference refinement P HOI tracking/contact teacher/refiner full per-sequence refiner baseline
SUGAR noisy kinematic prior → privileged refiner → dataset → tracker/generator P contact labels yes full closest end-to-end pipeline competitor
InterPrior full-reference expert → sparse-goal variational policy → RL P contact/keypoint/object goal yes paper/project only teacher + student competitor, not only student template
HumanVLA state teacher → vision/language student P object rearrangement yes full teacher/student external baseline
ComplexMimic imperfect MoCap → dual experts → difficulty-aware distillation P complex scene collision/contact distillation full noisy reference + complex scene competitor
ContactMimic reference keypoints + binary contacts → policy P explicit per-part contact command deployment command full contact contract / ablation source
WristMimic body/wrist reference + object outcome → finger policy P fingers not pose-supervised teacher only code pending finger-interface ablation source
InterReal augmented HOI reference → reward-learning tracker P hand contact augmentation teacher not verified reward-learning teacher competitor
ULTRA neural retargeting → tracker → dense-ref/sparse-goal/egocentric controller P G1 whole-body loco-manipulation yes paper/project; runnable release unverified direct tracker-to-autonomous-controller system competitor
CWI MoCap upper imitation + lower-body experts → whole-body teacher → bimanual-hand/command student P bimanual command, not full object HOI reference yes paper/project; runnable release unverified sparse deployment actor competitor
HumanoidMimicGen contact-rich demonstrations + planning → physical data → visuomotor policy P single/dual-arm contact skills data → policy paper/project; runnable release unverified physics data-engine/student competitor
VisualMimic low-level human keypoint tracker → egocentric visual keypoint policy P keypoint command, not dense HOI contact yes paper/project; runnable release unverified tracking-to-perception controller system competitor

接入与写作规则

  1. GRAIL-Min、A2-R、A2-G 都是 teacher candidates/baselines。 任何 adaptation 必须写清 body mapping、control space、reference visibility、object/contact condition 和 checkpoint interface;不准借用论文原名掩盖 interface change。
  2. SUGAR、InterPrior、HumanVLA 是完整 teacher/student 竞争系统。 即使无法在当前 runtime 直接复现,它们也必须进入 paper related work 和 claim boundary。
  3. physics-aware 不等于 P InterPhys/PAMotion 的 force/contact loss 说明 input 质量可能更好,但不能代替 physics rollout metrics。
  4. full 不等于项目已跑通。 每个 official code 仍须标为 native reproduction、adapted result 或 planned adaptation;运行证据只写在 Evidence Dashboard。
  5. controller paper 与 checkpoint contract 要分开引用。 原始 PULSE 报告 32D latent; 本项目 PULSE-X checkpoint 的 48D interface 是本地 checkpoint/code contract。两者不可 互相替代,也不可由“ProtoMotions 支持 SMPL-X”推出“存在可用 SMPL-X MaskedMimic”。