Network Ablation (Abl 4.0): Gap Terrain Instability (Raw Scan Critic) | Unitree Go2, Isaac Lab

Описание: Comparative experiment (Ablation 4.0) investigating the effect of feeding raw high-dimensional scan data to the Critic network. This video demonstrates the robot attempting to cross the gap terrain, revealing instability and foot-slipping behaviors compared to the encoded latent model (Abl 3.5).

📌 Experiment ID: Ablation 4.0 (Actor: Encoded Scan / Critic: Raw Scan) The goal of this ablation was to test if providing the Critic with the full, uncompressed 187-dimensional scandot data would improve value estimation. However, the results indicate that raw high-dimensional input introduces noise, making it difficult for the Critic to learn effective state values.

🕳️ Terrain Challenge: Gap
• Max Width: The gap width reaches up to 80cm, requiring the robot to perform a precise long-stride or jump.
• High Risk: Unlike flat ground, any misstep results in the robot falling into the void, leading to immediate episode termination.

⚙️ Network Architecture (Ablation 4.0):
• Actor: Proprioception + Encoded Terrain Latent (z_scan)
• Critic: Proprioception + Raw Privileged Obs + Raw Scandot (187D)

📝 Result Analysis:
• Instability: Unlike Ablation 3.5, the robot struggles to maintain stability. Notice how the left foot slips into the gap or drags near the edge during the crossing.
• Critic Inefficiency: The Critic struggled to filter meaningful features from the raw scan data. This noisy value estimation provided inaccurate gradients to the Actor.
• Local Minima: As a result, the Actor failed to learn clean movement patterns, falling into local minima (e.g., developing a habit of dragging feet) rather than anticipating the terrain.

💻 Code & Resources: github.com/yobel-sungkooklee/extreme-quadruped-parkour
#reinforcementlearning #robotics #IsaacLab #AblationStudy #simulation #unitreerobotics #UnitreeGo2