Seven Research Routes

Before selecting a route, use the 45-node Knowledge System if you need to resolve exact prerequisites. A route defines the research outcome; the knowledge graph defines the concept-level order and assessment.

English · 简体中文

This layer turns a broad interest into an executable research brief. Each route connects a question to prerequisite lessons, registered Pipelines, a concrete deliverable, metrics, a promotion gate, and an explicit evidence boundary.

A route organizes work; it does not upgrade the evidence level of any Pipeline. Check each linked Pipeline's status before making a claim.

Pick by research outcome

Direction	Build this	Core Pipelines
Foundation Models & VLA	Language-conditioned policy + adapter + ablation	Data → VLA → RFM
Manipulation & Imitation	Closed-loop manipulation baseline + failure taxonomy	Data → VLA → RL
Dexterity & Teleoperation	Retargeted motion + contact-aware grasp experiment	Retargeting → State → Grasping → Sim-to-Real
Navigation & Embodied Agents	State-aware agent loop + recovery report	State → Navigation → Reasoning
Humanoids & Locomotion	Motion protocol + safety and transfer gates	Locomotion → RL → Sim-to-Real
Perception & World Models	Uncertain state stream + predictive rollout	State → World model
Simulation, Data & Evaluation	Dataset datasheet + benchmark + promotion decision	Data → World model → Sim-to-Real

Inspect the machine-readable route contract:

python scripts/run_learning_path.py --list
python scripts/run_learning_path.py --show foundation-models-vla
python scripts/run_learning_path.py --validate

1. Foundation Models and VLA

Research question. How can language, vision, and robot state produce useful closed-loop actions?

• Prerequisites: deep learning, Transformers, datasets and training, evaluation
• Pipeline sequence: Simulation and Data → VLA Policy → RFM and Cross-embodiment
• Deliverable: a language-conditioned closed-loop policy, an adapter contract, and an ablation report
• Measure: task_success_rate, language_condition_gap, inference_latency_ms, adapter_coverage
• Promotion gate: beat the declared baseline under the same data and protocol, then pass action-schema and safety checks
• Evidence boundary: the RFM path is interface-tested; no competitive large-scale model reproduction is claimed

2. Manipulation and Imitation Learning

Research question. How can demonstrations become a robust manipulation policy?

• Prerequisites: kinematics and IK, control, datasets and training, evaluation
• Pipeline sequence: Simulation and Data → VLA Policy → RL Post-training
• Deliverable: a manipulation baseline with dataset diagnostics, closed-loop evaluation, and a failure taxonomy
• Measure: task_success_rate, selection_accuracy, collision_rate, inference_latency_ms
• Promotion gate: show a repeatable task gain across fixed seeds without increasing safety violations
• Evidence boundary: PushCube is a teaching-scale task, not evidence of general real-world manipulation

3. Dexterity, Retargeting, and Teleoperation

Research question. How can human motion be transferred to a robot hand without confusing geometry, contact, and task evidence?

• Prerequisites: SE(3), kinematics and IK, optimization, perception and sensors
• Pipeline sequence: Dexterous Retargeting → Perception and State → Dexterous Grasping and Fine Manipulation → Sim-to-Real
• Deliverable: a retargeted joint sequence plus a contact-aware grasping experiment, with geometry, temporal, collision, retention, task, and hardware evidence reported separately
• Measure: retargeting_error, joint_limit_violation_rate, latency_ms, grasp_success_rate, mean_lift_height_m, max_lateral_slip_m
• Promotion gate: pass geometry and temporal checks, then contact establishment, lift, retention, slip, and robustness gates; require a separate gate before hardware execution
• Evidence boundary: the committed contact-dynamics smoke validates one abstract MuJoCo grasp task; it does not establish in-hand reorientation, learned-policy quality, transfer to a production hand, or real hardware

4. Navigation and Embodied Agents

Research question. How should an agent estimate state, plan, act, recover, and replan over long horizons?

• Prerequisites: frames, control, sensing, systems and safety, evaluation
• Pipeline sequence: Perception and State → Navigation and Locomotion → Embodied Reasoning
• Deliverable: a state-aware navigation loop with typed plans, safety events, recovery behavior, and a scenario report
• Measure: localization_or_state_error, goal_success_rate, collision_or_fall_rate, recovery_success_rate, replan_count
• Promotion gate: meet scenario-level success, collision, and recovery thresholds under fixed maps and perturbations
• Evidence boundary: the runnable path is synthetic grid navigation, not a SLAM, mobile-manipulation, or legged benchmark reproduction

5. Humanoids and Locomotion

Research question. How can motion policies track commands, recover, and remain inside safety envelopes?

• Prerequisites: SE(3), kinematics, control, MuJoCo, systems and safety, evaluation
• Pipeline sequence: Navigation and Locomotion → RL Post-training → Sim-to-Real
• Deliverable: a locomotion protocol with tracking, perturbation, recovery, safety, and transfer gates
• Measure: path_or_velocity_tracking_error, collision_or_fall_rate, recovery_success_rate, sim_real_gap
• Promotion gate: pass motion-only simulation and safety regression before hardware-in-the-loop or robot trials
• Evidence boundary: no reproduced humanoid policy or locally validated humanoid hardware result is claimed

6. Perception and World Models

Research question. How can uncertain observations become state estimates and useful predictive rollouts?

• Prerequisites: frames, probability, perception and sensors, evaluation
• Pipeline sequence: Perception and State → World Model and Planning
• Deliverable: an uncertainty-aware state stream, a predictive rollout model, and calibration-to-planning error analysis
• Measure: calibration_reprojection_error_px, sensor_sync_skew_ms, uncertainty_calibration_error, multi_step_rollout_error, planned_task_success_rate
• Promotion gate: demonstrate calibrated uncertainty and improved downstream planning under the same observation protocol
• Evidence boundary: synthetic state and teaching-scale dynamics do not establish open-world visual prediction or real-sensor robustness

7. Simulation, Data, and Evaluation

Research question. How can experiments produce traceable data and evidence that survives comparison?

• Prerequisites: MuJoCo, datasets and training, systems and safety, evaluation
• Pipeline sequence: Simulation and Data → World Model and Planning → Sim-to-Real
• Deliverable: a versioned dataset datasheet, benchmark report, raw metrics, and an explicit deployment-promotion decision
• Measure: dataset_coverage, data_integrity_rate, task_success_rate, robustness_gap, sim_real_gap
• Promotion gate: retain protocol, seeds, data version, artifacts, negative results, and safety evidence before promotion
• Evidence boundary: repository audits verify committed contracts and artifacts, not external datasets or hardware deployment

Research loop

For every route: freeze the question → choose one baseline → run the minimum path → retain raw artifacts → analyze failures → change one variable → rerun fixed evaluation → promote only when the declared gate passes.

The canonical route data lives in learning_paths/manifest.json. Pipeline evidence status lives separately in pipelines/manifest.json.