abstract-robots

The abstract command and the robot station, on Abstract's robot API. It is built on abstract-client, the SDK (import atoms), which is all a robot with its own stack needs, and abstract-atoms installs both: Dream Labs' dreamlabs, dreamlabs-client and dream-robots.

uv tool install abstract-atoms  # the `abstract` command, on your PATH

abstract login | status | token | logout     your key, saved for this machine (readable by you alone)
abstract models                              robot models you can run, and what an hour of motion costs
abstract robot run --model A [--model B ...] --task "..." --robot my_robot:Arms [--dry-run] [--view]
                                             models on your robot: the first drives, the others compare
abstract datasets upload | list              your LeRobot datasets
abstract finetune create | list | get | events | cancel | models | download
abstract env init | check | eval | push | upload | grades     your own simulation environment
abstract skill [install]                     the skill that teaches an agent the API

Every command has --help with examples and --json where it prints data (data on stdout, progress on stderr), and exits 0 done, 1 failed (one line says why and what to run), 2 wrong usage, 130 stopped. Spending asks first (finetune create shows its quote; --yes for scripts, and it never asks without a terminal), and every file moved is checked by sha256.

abstract robot is the robot station, with Dream Labs' dreamcli robot commands for any number of models. robot run drives your robot through a driver class of yours with two methods: observe() returns the frames (one JPEG per camera) and both arms' state, and command(left, right) sends one row to each arm. The first --model drives the arms; every other one answers the same observations, and each prints its steps, rows, round trip and first pose side by side. --view serves a page on this machine that draws every model's planned grasp-site paths live, one colour each. --dry-run commands no arm, --robot still is a stand-in with no hardware, and it runs until Ctrl-C or --chunks: each chunk is billed. A station config (station.toml in ~/.abstract/station, or --config) holds the task, models, robot and chunks; flags override it.

Python 3.10 or later. It depends on abstract-client at its own version, since the two are released together, and on Python 3.10 on tomli. Tests: pytest apps/atoms.