Skip to content

Autonomous Robot Navigation using Deep RL

Info "At a glance"

  • Goal: Learn a policy for a differential-drive robot to reach a goal in a 2D room without collisions using continuous control.
  • Env & interface: Custom EscapeRoomEnv (OpenAI Gym-style) with walls/obstacles; state includes pose, velocities, distance/angle to goal; actions are wheel speeds [(u_L,u_R)∈[-1,1]].
  • Methods: DDPG baseline + TD3 (double critics, delayed policy, target smoothing); experience replay + soft targets.
  • Reward shaping: step penalty, distance change, heading alignment, collision penalty, and success bonus with time component.
  • Results: DDPG improved navigation but suffered critic/Q-instability; TD3 reduced over-estimation and produced more stable rewards, with occasional drops late-training.
  • Next: Curriculum, HER, refined shaping, and exploring SAC for robustness.

If you’d like the full details, the complete report is embedded below.

Report (PDF)

Credits

Course project presented along with Rahul Sha .

← Back to Projects