Public research log
Notes from the agent and robotics lab.
Deep dives on the systems we are building: Codex control loops, agent training infrastructure, simulation workflows, and the connective tissue that turns experiments into repeatable tools.
Current focus
Controllable agent runtimes, simulation workflows, and robot development environments.
Latest
Latest posts
Robot Policies: From π(a|o) to Foundation Models
A long-form research guide for generative-AI builders crossing into robot learning: policies, action spaces, training paradigms, locomotion, manipulation, deformables, frontiers, and the Unitree G1 MuJoCo code path.
Building a MuJoCo IDE for the Unitree G1
Cybernetic IDE puts Python control code, a Unitree G1 MuJoCo scene, and an embedded Robot Viewer in one development loop.
Turning Codex app-server into simulation workflows
A systems write-up on codex-control, builder/judge loops, and how our Workflows feature turns a phone prompt into a live Isaac Sim scene.
Field notes
Get the next research note.
Occasional dispatches on agent control loops, simulator infrastructure, robot development systems, and the practical details behind what ships.
Stack map
The packages in the first essay.
Runtime control
codex-control
Typed async access to Codex app-server threads, turns, fork, steer, interrupt, rollback, traces, budgets, and transports.
Working proof
research-swarm
A local-first research harness that uses codex-control as its Codex backend while Python owns tools, archives, citations, and traces.
Model wire
codex-proxy
Responses API translation, SSE serving, per-rollout record capture, token logprobs, and drop-in inference for trained policies.
Task contract
codex-env
Pure stdlib task and verifier definitions: prompts, public/private workspaces, reward results, and grading batteries.
Runtime conductor
codex-orchestrate
Job runners, workspace staging, snapshots, verifier execution, Codex session pools, RLVR episodes, and rollout groups.
Learning loop
codex-train
GRPO/DAPO math, RolloutProvider-driven training, Worldlines datums, checkpoint manifests, and policy update hooks.