Hydraulic excavator — learning to dig
A 50 t class excavator learning to dig in VeloxSim. It practises in our NDEM granular simulation, where the soil pushes back on the bucket and the hydraulics. After 220 practice digs, about 7 minutes of training on one GPU, it fills the bucket with 4.3 t instead of 2.6 t, on a bed it never practised on.
A small neural network learns a correction to the excavator's fixed dig plan by reinforcement learning (TD3), from each dig's score alone: the load in the bucket, less penalties for time with the hydraulics at relief pressure and for the energy used. It is told only what a machine can measure, such as cylinder strokes and pressures, not the height of the ground. Untrained, its output is close to zero, so the machine simply runs the plan. Every practice dig is a full dig simulated in NDEM, and 220 of them took about 7 minutes of training on one NVIDIA RTX 5090.
The bed in the clip is 0.44 m lower than the plan expects, and lower than any ground the network practised on. The plan alone lifts 2.6 t from it. After 120 practice digs the network lifts 2.1 t, less than the plan; after 220 it lifts 4.3 t.
The grid replays all twelve of the run's saved checkpoints side by side on that same bed, one every 20 practice digs, fewest first. Progress is not steady: at 40 practice digs the network lifted 0.9 t, and at 80 it lifted 4.3 t but held the hydraulics at relief for more than 3 seconds. The heaviest single dig, 4.4 t, came at 200, but the 220 dig has the best score of all twelve: it used 7% less energy than the 200 dig and spent less time at relief, and at 4.2 t per MJ it moved the most material for the energy of any checkpoint.
The dig shots play at 2.5× speed, with 80 mm grains, rendered in Isaac Sim. The load on screen is a live count: it reads high while grain is pushed into the bucket and settles as loose grain falls off on the lift. The figures above are each dig's result. The grains use generic material parameters and have not been calibrated.
Simulated with NDEM, our real-time granular engine.
Robotics team doing granular simulation?
If you're building robots that dig, load, or move bulk material, send us a geometry or a scenario — we'll show you what it looks like in NDEM: real-time and calibrated to real materials.