On-edge self-improvement · RL-environment pre-training

AGI for robots that learn & improve on-edge.

Two systems, one robot: an on-board governor that makes your robot faster and better by itself after it ships — and RL environments that pre-train it to cross the reality gap on the first try.

loading G1…

Measured against current RL pipelines

Cost-effective. GPU-efficient. Self-improving.

Time to fine-tune · vs. training data

Current RLCadenzaTRAINING DATA →FINE-TUNE TIME

The more data you train on, the further Cadenza pulls ahead, fine-tuning in a fraction of the time current RL systems need.

+80%Accuracypre-training vs. current RL sequences on closed-loop manipulation
2.1×Faster, on its ownthe arm self-improves on-edge after deploy — no retraining

01 · The systems

Two systems. One self-improving robot.

Pre-train a policy on RL environments that actually cross the reality gap — then let it keep getting better on the robot itself, on-edge, long after it ships.

01
On the edge

Robots that scale & learn by themselves

An on-board governor watches the robot succeed and makes it faster and better — by itself, after it ships. No retraining, no cloud.

  • Self-improves in place. Tiny, committed speed-ups per task while the robot runs.
  • Safety-gated. Keeps a change only if the outcome holds; reverts the instant it wouldn't.
  • Long-horizon. Microstep checkpoints turn long tasks into sub-goals it solves one by one.
02
Pre-training

RL environments that cross the gap

Novel, GPU-efficient RL environments that auto-calibrate physics against real telemetry, so policies are deploy-ready before they touch hardware.

  • Sim-to-real fidelity. Domain randomization tuned to your captured robot telemetry.
  • Linear scaling. 100k deterministic instances per GPU partition, one seed.
  • Any robot. URDF, MJCF, or a Cadenza spec — arm, hand, or quadruped.

02 · Pre-training

Pre-train policies on RL environments in a dozen lines.

Name the robot and the scene, and Cadenza builds the RL environment, auto-calibrates physics against real telemetry, and scales to hundreds of thousands of parallel instances. The policy you pre-train here is the same arm that keeps improving itself on-edge below.

  • Declarative tasks. Reward functions defined inline and versioned, not glued code.
  • Sim-to-real fidelity. Domain randomization tuned to your captured robot telemetry.
  • Linear scaling. 100k deterministic instances per GPU partition, one seed.
build_env.pypython
1import cadenza as cz
2 
3# Build an RL environment for the Cadenza arm on Cadenza Lab.
4# Physics auto-calibrates against captured telemetry.
5env = cz.Environment(
6 robot=cz.robots.Arm(dof=6),
7 backend="mujoco",
8 scene="lab/stacked_blocks",
9 fidelity="sim2real", # domain-randomized, contact-rich
10)
11 
12@env.task("pick_place")
13def reward(state, action):
14 placed = cz.metrics.object_at(state, target="left")
15 stable = cz.metrics.tower_intact(state)
16 effort = cz.metrics.energy(action)
17 return 4.0 * placed + 2.0 * stable - 0.01 * effort
18 
19# Scale to 100k parallel instances on one GPU partition.
20fleet = env.parallelize(n=100_000, seed=7)
21 
22policy = cz.train(
23 fleet,
24 algo=cz.algos.PPO(clip=0.2, gae_lambda=0.95),
25 steps=2_000_000,
26)
27 
28policy.export("arm_pick_place.cadenza") # ships with the on-edge governor

03 · On the edge

Self-improvement that ships with the robot.

Give the arm a goal in plain English. It resolves the command into the action library, tokenizes it into microstep checkpoints, and runs the task on a delicate three-block tower. Then the on-board governor makes it faster on its own, rep over rep — keeping every speed-up that leaves the tower standing, backing off the instant one wouldn't. No retraining, no cloud.

booting arm…
Cadenza arm6-axis · on-edge
idle
approachdescendgraspliftcarrylowerreleaseretract
rep 0speed 1.00×tower shift 0.0mm / 15status
Layer 1

SemanticLayer

natural language → action library

Layer 2

MicrostepTokens

per-phase checkpoints to solve

megan-tk

On-edge governor

speeds up; reverts if the tower shifts

drive_arm.pypython
1import robogpt
2 
3arm = robogpt.connect("cadenza-arm") # 6-axis, on-edge governor
4plan = arm.understand(PROMPT) # NL → action library
5plan = robogpt.microsteps(plan) # per-phase checkpoints
6 
7for ep in arm.run_forever(plan): # self-improves in place
8 if ep.clean and ep.faster:
9 arm.commit(ep.speeds) # keep the speed-up
10 else:
11 arm.revert() # tower shifted → back off

04 · Who it's for

Built for teams shipping self-improving robots.

01

For inference startups

A rebuilt rollout engine reclaims 70% more GPU headroom, so you can serve more robot policies per partition without buying more silicon.

02

For robotics developers

Everything to take a policy from sim to a shipping robot: RL environments, data, evals, and an on-edge governor that keeps it improving — in one layer.

03

GPU-efficient by design

Cost-effective rollouts that scale linearly to 100k deterministic instances per GPU partition, one seed.

04

Reliable across the gap

Contact-rich physics auto-calibrated against captured telemetry, so policies cross the reality gap on the first deploy.

05

Any robot description

URDF, MJCF, or a Cadenza robot spec. Bring an arm, a hand, or a quadruped, and the systems adapt to your hardware.

06

Self-improving in the field

Export one artifact that runs in sim and on hardware, with the on-edge governor making it faster and better after it ships.

Ship robots that
improve themselves.

Two systems for self-improving robots: RL environments to pre-train them, and an on-edge governor that keeps them getting better in the field. Request access and we'll stand it up with you.