- Guideline: installation, hardware-free demos and complete YAM startup commands.
- Configuration guide: config layout, field meanings and shared control profiles.
- Viewer tutorial: rollout controls, camera views and experiment labels.
- Experiment workflow: persistent services, normal/experiment modes and saved evidence.
- YAM collection: the original-style collection GUI with ManiMux follower control.
- Architecture: policy, adapter, strategy, executor and robot boundaries.
- Agent guide: XPolicyLab-only model integration, legacy migration and validation rules.
- XPolicyLab: model adapters and model-side sampling behind the shared policy interface. See the integration runbook.
- PRM-as-a-Judge: offline model-based evaluation of recorded videos. See the evaluation guide.
Both are version-pinned submodules. They are platform components, not additional policies or inference strategies in the support counts below.
- Teleop: available through the YAM collection GUI, with ManiMux follower control.
- UMI: planned collection integration; no runnable collector entry point yet.
- DAgger: planned human-intervention / corrective-data collection; no runnable collector entry point yet.
The Collection badge groups current and planned collection modes. The LLM-judge badge refers to the offline model-judge integration through PRM-as-a-Judge, not an additional online controller.
Use each runbook's checkpoint, environment and action contract together. An available adapter does not mean every checkpoint or inference-method combination has passed a real-robot trial.
- Pi05 / OpenPI, including paired put-bottles joint / joint+EE 30k configurations.
- MolmoAct2 · ABC.
- GR00T N1.7 · LingBot-VLA2.
- Xiaomi XR-1 · OpenWAM.
- SAPolicy · Shared XPolicyLab bridge.
- Offline / simulation paths: Cosmos3, Isaac 0.5, ManiUniCon.
The README badges count integration coverage, not task success, hardware validation of every checkpoint, or support for every policy × embodiment × inference combination.
- 10 policy integrations: eight model families have YAM deployment configurations: Pi05, MolmoAct2, ABC, GR00T, LingBot-VLA2, Xiaomi XR-1, OpenWAM and SAPolicy. Cosmos3 and Isaac 0.5 add two model-only / offline paths, not two more YAM-ready policies. Checkpoint variants, the generic XPolicyLab bridge and the collection leader policy are not counted separately.
- 1 real + 1 simulation embodiment integration: the dual-YAM driver and the ManiUniCon/Meshcat simulation driver. Mock hardware is excluded. A model's DROID or LIBERO checkpoint does not itself establish a ManiMux robot-driver integration.
- 8 inference modes: seven built-in strategies—ManiMux, RTC, ACT temporal ensembling, AAC, PAINT, AutoHorizon and DVAC—plus serial prefix execution. Serial is a scheduling mode of the ManiMux strategy, not an eighth registered strategy. Direct, Smooth and MPC are executors and are not counted as inference methods.
Sources: model configurations, robot factories, strategy registry and serial execution. UMI and DAgger are collection roadmap items and do not contribute to these implementation counts.
Method documents distinguish the upstream method, ManiMux adaptation and validation evidence. Choose a method supported by the policy backend; these are not interchangeable sampler hooks.
- ManiMux asynchronous scheduling · Serial prefix execution.
- RTC contract · Chunk-step semantics.
- ACT temporal ensembling.
- AAC · Pi05 AAC.
- Pi05 PAINT.
- Pi05 AutoHorizon · Pi05 DVAC.
- Braking execution · Independent IK.
- Experiment design: controlled comparisons and evaluation planning.
- PRM-as-a-Judge: offline video manifests, judge setup and reports.
- Reference-layout overlay: reproduce an experiment's initial scene.
- YAM training pipeline · Pi05 joint+EE training.
- CAN setup: YAM-specific hardware configuration.