Object Capture 2D — Sensor-Guided Dwell Grasp

Training

Sim speed 1.0×

Obstacle density 7

Capture telemetry

Reward

0.00

Success rate

Captures

Collisions

Dwell progress

Exploration ε

100%

+1.0 capture-0.3 collision

Rover body / heading Sensor rays (5, front arc) Obstacle (collision = -0.3) Target + capture-dwell ring

How this differs from the 3D model

The 3D Object Capture environment is a multi-agent potential field: two dozen massless points glide toward the nearest goal orb and steer off obstacles by an inverse-square repulsion term, "capturing" the instant they pass within a small radius. This 2D companion opens a single, physically-driven rover instead:

drive:    v̇ = a_cmd,  θ̇ = ω_cmd            (differential-drive kinematics, not a massless point)
sensing:  d_i = raycast(θ + β_i), i = 1..5    (5 fixed-angle rays, no omniscient obstacle list)
capture:  hold inside r_cap AND |v| < v_dwell for T_dwell seconds → capture event

The rover never knows obstacle positions directly — it only sees five ray distances fired from its own nose, exactly like a real bump/IR sensor ring. Steering blends target bearing with the nearest ray's push-back.
Reaching the target isn't enough: the rover must actively brake and dwell inside the capture ring for a fixed time before the object counts as captured — a graze-and-leave gets no reward, matching how a real gripper needs to stop and close, not just touch.
"Training" is expressed as a decaying exploration term ε that shrinks episode over episode, injecting steering noise early on and converging to a clean approach later — visible directly in the reward curve's smoothing trend, the same qualitative signature the 3D dashboard shows via its policy-loss curve, produced here by a completely different rover/sensor model.

Obstacle density is adjustable — more obstacles make the raycast-avoidance term dominate longer before the rover can commit to the final dwell approach.

Rover Obstacle Target

Top-down arena — raycast sensors

Approaching

Reward curve