Demonstration Trajectory

A hands-on method where a human expert physically guides a robot's end-effector or joints through the desired task motions. The robot's proprioceptive sensors (encoders, torque sensors) record the resulting state-action pairs.

• Primary Use: Ideal for collecting precise, low-level control demonstrations for manipulation tasks (e.g., pick-and-place, assembly).
• Advantages: Generates kinematically feasible, smooth trajectories directly in the robot's own state space, avoiding correspondence problems.
• Limitations: Can be physically demanding; demonstrations are constrained by the robot's own dynamics and range of motion.

The expert controls the robot remotely using an input device (e.g., joystick, space mouse, VR controllers, or a master robotic arm). The commands and resulting robot states are logged to form the demonstration trajectory.

• Primary Use: Essential for collecting data for mobile robots, drones, or complex manipulation in hazardous or remote environments.
• Interface Types: Includes direct mapping (joystick to velocity) and supervisory control (clicking waypoints).
• Advantages: Allows collection of demonstrations for tasks impossible or unsafe for direct physical guidance.
• Challenge: Can suffer from time delays and requires the operator to learn the mapping between interface and robot motion.

Demonstrations are captured by observing an expert (often a human) performing the task from an external viewpoint, using cameras or motion capture systems. This yields state-only trajectories (e.g., video frames, object poses).

• Primary Use: Enables leveraging abundant human video data (Imitation Learning from Observations) or demonstrations from a different embodiment.
• Core Challenge: The correspondence problem—the agent must learn to map observations from the demonstrator's viewpoint and embodiment to its own action space.
• Techniques: Often solved with domain adaptation or learning viewpoint-invariant representations.

Demonstrations are generated programmatically within a physics simulator. The 'expert' can be a scripted policy, an optimized trajectory (e.g., from trajectory optimization), or even another trained AI agent.

• Primary Use: Enables scalable, automated collection of massive, perfectly labeled datasets for sim-to-real transfer pipelines.
• Advantages: Provides infinite, risk-free data; allows access to privileged information (full state); easy to add noise for robustness.
• Key Consideration: The reality gap means policies trained on simulated demonstrations may fail on physical hardware without domain randomization.

A trajectory is not just a list of actions; it's a temporally aligned record of the system state. This often requires fusing data from multiple sensors to reconstruct an accurate state estimate.

• Common Sensors: Inertial Measurement Units (IMUs), joint encoders, force-torque sensors, external motion capture (OptiTrack, Vicon), and on-board cameras.
• Process: Raw sensor readings are filtered (e.g., using a Kalman Filter) to produce a coherent state vector (position, orientation, velocity) for each timestep in the trajectory.
• Importance: Noisy or inaccurate state estimates directly degrade the quality of the learned policy.

Raw recorded data is rarely training-ready. Critical post-processing steps structure the demonstration into a usable dataset.

• Temporal Alignment: Synchronizing clocks across all sensors and the control system.
• Action Inference: In state-only demonstrations, actions must be inferred via inverse dynamics models or by differentiating state.
• Frame Transformation: Converting all poses and observations into a consistent coordinate frame (e.g., the robot's base frame).
• Trajectory Segmentation: Isolating successful task episodes from continuous recordings and labeling them with task identifiers or success signals.