Askesis
Physical skill, learned through practice.
Askesis is our training-data product for robotics and space teams: demonstrations of off-world construction work, and the environments they were recorded in, in formats your tooling already reads.
Askesis is the Stoic word for training through practice. A pianist or a veteran excavator operator learns physical skill by doing, not by being told. Machines learn it the same way.
In development
Askesis isn't on sale yet. We're looking for design partners to shape what goes into it.
What exists today
Observation and action pairs, with provenance on every session.
Session logs
JSON Lines at 10 Hz: pose, velocities, boom and bucket state, bucket-edge position, load, task progress, the terrain heights around the machine, per-step soil forces, wheel sinkage and slip, and the four control inputs. Robot-camera frames at 2 Hz.
Provenance
Every session header records the controller (human, scripted or learned), the world, the terrain source, the control mode, the lighting, and the soil parameters with their citations. The schema only ever grows, so old data stays readable.
LeRobot export
A converter writes sessions as a Hugging Face LeRobot dataset, one session per episode: a 20-value state, an 11 × 11 terrain patch, the robot camera image and a 4-value action. It loads with LeRobot's own training scripts.
Job scoring
Every finished job is scored the same way whoever did it: depth against target, evenness, soil moved and energy used. That lets you filter demonstrations by quality, not just by completion.
Where we are, honestly
Scripted-expert data proves the pipeline. Human data is next.
- The policy we've trained and shipped learned from scripted-expert data only. It digs the trench to spec on 47 of 50 held-out starts. See research.
- Human sessions so far: 34, from one player, used in tests, not in the shipped policy. Collecting player data at scale waits on consent terms and an opt-in upload.
- Policies today read exact machine state, not camera images. Camera-based policies are on the roadmap.
- One machine (an excavator), two worlds, two jobs. More machines and jobs follow the order missions will need them: site prep, berms and landing pads, regolith hauling, module assembly.
Roadmap
What Askesis grows into.
Diverse human strategies
Many players solving the same job different ways: the variety one scripted expert can't give.
Randomized conditions
Soil properties, sensor noise and lighting varied across sessions, standard practice for sim-to-real transfer.
High-fidelity twin
Sessions replayed in Project Chrono's deformable-terrain simulator to measure where the fast model drifts.
Planning data
Supervised-autonomy plans paired with outcomes, to train how a planner should command a machine like this.
Custom environments
Your site, your machine, your soil: built to order and validated against your measurements.
Skill models
Later, trained skill models licensed per machine, starting with Earth construction and mining.