Project Triskelion features a 3-legged hopping robot trained via Deep Reinforcement Learning to navigate low-gravity terrain[cite: 3]. Built completely end-to-end—from custom CAD to URDF pipelines and PPO controllers—the system is designed for fault-tolerant planetary rover applications. The architecture is designed to handle varying gravitational environments, scaling from Mars (0.3g) down to the Moon (0.1g) and Ceres (0.029g).
This project extends prior discrete-jump reinforcement learning research (such as ETH Zürich's SpaceHopper) by achieving goal-directed hopping navigation under low-gravity constraints.
phaseD_demo_static_fwd3m.mp4
The following demonstration showcases the final Phase D policy. The robot successfully hops from a random spawn point to an arbitrary target located 2 to 5 meters away.
- Environment: Moon gravity (−0.981 m/s²).
- Control: The 120°-symmetric body is completely omnidirectional; the robot does not need to face the target to navigate effectively.
https://github.com/Kairat11/triskelion-simulation-demo/blob/main/phaseD_demo_static_fwd3m.mp4
The final policy (checkpoints/phaseD_v22_latest_79.9M.pt) was trained using a CleanRL-style PPO architecture within the Newton physics engine. The reward mechanism successfully drives navigation via distance-reduction (pos_delta) rather than strict velocity-tracking.
Omnidirectional navigation is fully solved: The policy reaches within 0.5 m of arbitrary 2–5 m targets with exceptional reliability and zero falls.
*Evaluated over 500 steps per environment at Moon gravity.
| Target Direction | Arrival Rate (< 0.5m) | Fall Rate | Median Closest Approach |
|---|---|---|---|
| Forward (3 m, +x) | 96.1%[cite: 2] | 0.0%[cite: 2] | 0.17 m[cite: 2] |
| Diagonal (2, 2) | 98.4%[cite: 2] | 0.0%[cite: 2] | 0.15 m[cite: 2] |
| Right (3 m, +y) | 100.0%[cite: 2] | 0.0%[cite: 2] | 0.16 m[cite: 2] |
| Left (3 m, -y) | 99.2%[cite: 2] | 0.0%[cite: 2] | 0.14 m[cite: 2] |
| Back (3 m, -x) | 97.7%[cite: 2] | 0.0%[cite: 2] | 0.16 m[cite: 2] |
| Near (1 m, +x) | 100.0%[cite: 2] | 0.0%[cite: 2] | 0.10 m[cite: 2] |
Note on Directional Bias: Lateral directions (left/right/back) were specifically tested to ensure the 120°-symmetric gait did not suffer from a weak axis; the data confirms no directional bias exists.
The simulation and training pipeline requires no subtask decomposition, utilizing a single policy to handle locomotion, hopping, and procedural terrain traversal.
- Physics & Kinematics:
- Direct OnShape CAD to URDF preprocessing, ensuring visual geometry matches physics colliders.
- Simulated utilizing the Newton physics engine for high-fidelity rigid-body contact dynamics.
- Reinforcement Learning:
- CleanRL-style PPO (256/256 MLP actor and critic).
- 48-dimensional observation space mapping to 9-dimensional joint offsets.
- Procedural Environment:
- Features a 12-block heightmap curriculum mixing wave and grid difficulties.
- Supports multi-gravity validation across regimes spanning a 10× factor.
While the navigation policy is highly robust, the current gait operates as a continuous hopper (airborne approximately 92% of the time). The agent reliably reaches the target spatial location, but success is currently defined by arrival rather than executing a full multi-foot parked stance, which the current hopping physics actively fight against.
Future development phases include executing gravity sweeps for asteroid environments (Ceres, -0.28 m/s²) and implementing stage 2.5 air-righting for cat-twist flight recovery.