Humanoid Imitation & Control: visual research map

GMR, BeyondMimic, SONIC, GRAIL, VIRAL, VisualMimic, Psi0, SIMPLE · architecture screenshots + synthesis · 2026-06-23

humanoidPPODAggerSONIC latentGMRsim-to-realVLAhardware

1 · Big picture

One-line takeaway. These papers are not all training the same thing: GMR retargets, BeyondMimic/SONIC train physics trackers, GRAIL/VIRAL/VisualMimic train privileged teachers and visual students, and Psi0 is mostly offline VLA-style behavior cloning / flow matching.

The figure below is the cleanest way to keep the stack straight.

Original synthesis diagram. The main pipeline is human/generated data -> retargeting/reconstruction -> privileged physics teacher -> compact representation -> visual/deployable policy.

2 · Training methods matrix

The overloaded words are the source of confusion. PPO usually trains the controller or teacher; DAgger trains the student on states it actually visits; Psi0 mostly uses offline supervised training rather than online DAgger.

Original matrix: which system uses PPO, DAgger, offline BC/flow, and latent/token losses.

3 · Paper screenshots: pipeline and architecture figures

These are cropped screenshots from the papers, included here because the actual architecture figures make the differences much easier to see than prose.

GRAIL Fig. 2 screenshot: asset-conditioned 4D HOI generation. The system starts from a specified 3D asset/scene, generates a reference HOI video, then reconstructs human and object trajectories.

GRAIL Fig. 3 screenshot: complementary controller adaptation. This is the key SONIC-based architecture: reference motion enters the frozen controller, while the object-aware adaptor injects a latent residual and hand primitive.

VIRAL Fig. 2 screenshot: privileged PPO teacher plus RGB/proprio student. Student training mixes DAgger and BC while targeting the teacher WBC command/action.

VisualMimic Fig. 2 screenshot: two-stage hierarchy. First a general keypoint tracker is distilled; then task-specific keypoint generators are trained/distilled.

Psi0 Fig. 2 screenshot: VLM pretraining, action expert post-training, and real-time deployment with lower-body control.

Psi0 Fig. 3 screenshot: MM-DiT action expert compared with a naive DiT head.

4 · GRAIL + SONIC latent adaptor

GRAIL is easiest to understand as a latent-control wrapper around a frozen SONIC controller. The generated trajectory provides q_ref; the frozen SONIC encoder maps it into a base latent; PPO only trains an adaptor that adds an object-aware residual.

Why not directly output joints? Direct joint PPO would throw away SONIC's learned whole-body prior. A latent residual keeps the action near a natural/balanced humanoid motion manifold while still allowing grasp/contact corrections.

Original architecture diagram of the GRAIL teacher and visual-distillation stages.

5 · Distillation patterns: VIRAL vs VisualMimic vs GRAIL

All three systems have a teacher-student flavor, but the student target is different: WBC command in VIRAL, sparse 18D keypoints in VisualMimic, and SONIC latent/action interface in GRAIL.

Original comparison diagram. The biggest practical difference is whether the teacher can be queried online for DAgger and what representation the student is trained to imitate.

6 · Representation choices

The compact representation is the design bottleneck. It determines the difficulty of the tracker, the visual student, and the sim-to-real transfer.

Original spectrum: full q_ref, sparse keypoints, learned SONIC tokens, WBC commands, and VLA action chunks.

7 · Hardware / simulator decision map

The short version: RT cores matter for Isaac RGB rendering, not for every learning algorithm. H100/H200/A100 are fine for offline model training on saved data; RTX/L40/L40S/4090/5090 are the practical choices for Isaac Sim visual rollout/rendering.

Original hardware decision map separating rendering, physics simulation, PPO tracking, and offline VLA training.

8 · Method summary table

System	Main role	Training method	Representation	Best mental model
GMR	Retargeting	No learning	Robot reference motion `q_ref`	Human motion -> robot kinematic reference
BeyondMimic	Motion tracker	PPO	`q_ref` / joints	Learn closed-loop tracking of GMR outputs
SONIC	Universal controller prior	Large-scale PPO + latent losses	FSQ latent token -> decoder	Motion tracking foundation controller
GRAIL	Generated HOI to real humanoid	PPO latent adaptor + visual distillation	SONIC latent residual/action	Generated `q_ref` steered through frozen SONIC
VIRAL	Visual sim-to-real	PPO teacher + DAgger/BC student	WBC command/action	Privileged teacher labels RGB student rollouts
VisualMimic	Visual humanoid manipulation	PPO teachers + DAgger students	18D keypoint command	Hierarchical keypoint interface
Psi0	Humanoid VLA	Offline BC / token prediction / flow matching	Action chunks	Large-scale offline VLA with real-time chunking
SIMPLE	Simulator/data layer	Not a policy	MuJoCo physics + Isaac rendering	Generate/evaluate RGB datasets

9 · Reproduction recommendation

1. Start with GMR
   human motion -> robot q_ref

2. Train a BeyondMimic-style PPO tracker
   q_ref + proprio -> joint action

3. Add a compact student interface
   VisualMimic-style 18D keypoints, or GRAIL-style latent if SONIC is available

4. Add visual distillation
   privileged teacher -> RGB/depth student
   use DAgger if the teacher can be queried online

5. Split rendering from learning
   RTX/L40S workers for Isaac RGB; A100/H100/H200 for offline model training

Practical ranking. Fastest start: GMR + BeyondMimic. Best public visual-distillation reference: VIRAL. Best sparse abstraction: VisualMimic. Most powerful latent route if SONIC is available: GRAIL. Hardest to reproduce from scratch: SONIC.

10 · Code status and references

Project	Code / page	Checked local HEAD
GRAIL	NVlabs/GRAIL	`b9ce9fdefe6f33c378b40cd5bba693bf7bd77fd5`
VIRAL	NVlabs/GR00T-VisualSim2Real	`92bf0863d4a9b6ee29849736152b7769bd45c49c`
VisualMimic	visualmimic/VisualMimic	`57643ccb2eea84ade58071ddab74d8a75387071c`
Psi0	physical-superintelligence-lab/Psi0	`c7bdd8421f517f003dc51bafb8a8d4ab64b3fb73`
SIMPLE	physical-superintelligence-lab/SIMPLE	`d010f0f537af97a0376168d00e7e27e159226b07`
BeyondMimic tracker code	HybridRobotics/whole_body_tracking	`cd65172032893724b445448818c34165846d847d`

GRAIL paper: arXiv 2606.05160
Psi0 paper: arXiv 2603.12263
VIRAL site: viral-humanoid.github.io
VisualMimic site: visualmimic.github.io
GMR / Retargeting Matters: project page

Generated for the humanoid research hub · visuals include original synthesis diagrams plus cropped paper figure screenshots · 2026-06-23