Record 100 demonstrations on a robot arm and train a policy: will it learn to go around the cup and pick up the box? What about a different table, a different instruction, a different robot? This series has just two notes. The first covers learning actions from demonstrations: why "taking the average" drives into the obstacle, and the trade-offs in action chunking and data. The second covers turning a vision-language model into a robot policy (VLA), and carrying what was learned in simulation over to a real robot.

Draft, pending review. The site owner has not reviewed this note yet; facts, emphasis and wording may change. Check the primary sources linked in the text before relying on it.

How the two notes divide the work

Note What it answers Hand-worked example After reading you should be able to explain
Learning actions from demonstrations Which extra problems behavior cloning meets on robots; what generative action heads and action chunking each fix; where the data comes from Left or right around an obstacle: MSE picks 0 and hits the cup; flow matching with only 1 step also falls back to 0 and needs several steps to split to either side; jitter and reaction delay for chunk lengths from 1 to 50 Why current policies output a stretch of actions at once, and why with diffusion or flow matching
VLA models and sim-to-real How a VLM becomes a robot policy; why humanoids split into fast and slow layers; how RL is added on top of a VLA; how to evaluate; how to close the gap from simulation to the real robot One 50-step × 7-D chunk: about 3.5 s with token-by-token decoding, about 70 ms with a flow-matching expert (teaching numbers); pushing a puck when friction is uncertain Where each VLA's actions come out and why; which models are open; what humanoid companies have officially said

An action travels from data to motors roughly through these steps; the two notes cover the first and second halves:

Demonstrations: teleoperation, hand-held grippers, first-person human video
→ Policy: images + instruction + robot state → a stretch of actions (action chunk)
→ Action head: regression / discrete tokens / diffusion or flow matching
→ Execution: sent to the controller at 20–50 Hz, re-inferred after part of the chunk
→ Evaluation: simulation benchmarks, double-blind real-robot comparisons
→ Improvement: RL on the robot's own experience; randomize in simulation, then deploy

What to read first

Read first Why
Imitation and rewards Behavior cloning, compounding error, DAgger, causal confusion; the first note assumes you have read it
Model-based RL "Averaging two paths into a wall" and MPC executing only the first action; both examples in the first note continue from here
Qwen-Image: flow matching The flow-matching objective and a 1-D example; the first note keeps its notation
LLaVA and DeepSeek-VL How images enter a language model; the second note assumes you know the basic VLM structure
Actor–Critic and GAE, How much should a policy change? (PPO), Offline RL and OPE Used in the second note for RL on top of VLAs and for training controllers in simulation

Optional: How a 3D point becomes a pixel (camera geometry, helpful for multi-view inputs); World models as simulators, evaluators and policies for robots and cars (not repeated here).

State as of October 2026

Each item below is as of 2026-10-15, with its source in the same sentence. "Self-reported" means the company published the number itself and nobody has checked it independently.

  • The usual VLA recipe: a pretrained VLM plus a continuous "action expert" (flow matching or diffusion) that outputs about 50 steps at a time, at up to 50 Hz. For example, π0 is PaliGemma 3B plus a 300M-parameter expert, and π0.7 is Gemma 3 4B plus an 860M-parameter expert; GR00T N1.x and SmolVLA follow the same pattern.
  • Humanoids split into fast and slow layers: Figure's Helix pairs a 7B VLM at 7–9 Hz with an 80M-parameter policy at 200 Hz, and Helix 02 adds a 1 kHz whole-body controller (self-reported); GR00T and Gemini Robotics 2 (2026-07-30) use the same split.
  • RL now runs on VLAs: Physical Intelligence's π*0.6 / Recap (2025-11) and RL tokens (2026-03) both report gains in throughput or speed (self-reported).
  • The data bet is first-person human video: NVIDIA's EgoScale uses 20,854 hours; Figure's Index claims more than 16 million videos; Generalist's GEN-0 claims 270,000 hours of manipulation data, growing by about 10,000 hours a week and collected by "1,000s of data collection devices and robots", without saying how much came from people versus robots (all self-reported).
  • Video world models are entering the policy itself: for example the 14B DreamZero (2026-02); NVIDIA's GR00T N2, previewed in March 2026, is based on it, and no official release was found as of 2026-10-15. Details are in the Frontier block's world-model note.
  • Openness is split: openpi (π0, π0-FAST, π0.5), GR00T N1.x, OpenVLA, Octo and SmolVLA have open weights; Gemini Robotics VLAs, Figure Helix and π0.6 / π0.7 do not. PI opened a hosted API beta on 2026-10-13 (PI blog). Licence details are in the second note's table.
  • Simulation is consolidating on GPU parallelism: Newton went to the Linux Foundation in September 2025, and version 1.0 was released on 2026-03-10; NVIDIA's press release of 2026-03-16 says Isaac Lab 3.0 (early access) is built on Newton 1.0 plus PhysX; MuJoCo Warp is also part of the Newton project.
  • Evaluation is moving to real robots: RoboArena runs double-blind pairwise real-robot comparisons; OpenVLA-OFT reports an average of 97.1% on LIBERO's 4 commonly used evaluation suites, and our judgement is that this simulation benchmark is near its ceiling.
  • Humanoid deployments are real but small: BMW confirms that one Figure 02 helped build more than 30,000 X3s; Tesla's update letter of 2026-07-22 says first-generation Optimus lines are being installed "in anticipation of production in 2026", with initial builds used first to collect training data. A company-by-company list based on official statements is at the end of the second note.

Sidewalk delivery robots

Delivery robots also run in the real world; of the three below, Starship has delivered the most and says it has completed 10 million deliveries. This section relies mainly on the companies' own material (all numbers self-reported) and keeps their wording:

  • Coco Robotics: its About page (read 2026-10-15) lists 500,000 deliveries and 1,000 robots produced (written as 500K and 1K), and says the two founders started the project as students at UCLA; press releases from June 2025 and 2026-07-15 (a reprint of the PR Newswire release) both say "over 500,000", unchanged for more than a year. Partners include Uber Eats (Los Angeles, from August 2024, per The Robot Report), Wolt in Finland, and Deliveroo in the UK, announced 2026-10-05. One inconsistency is worth keeping in view: on the About page, the CEO's bio says the company builds "human-piloted" sidewalk robots, while the rest of that page and press releases since 2025 call them "autonomous"; the Coco 2 page describes end-to-end neural networks with human oversight and a "Pilot Terminal" that steps in when needed. Coco has not published how much driving is autonomous versus remotely piloted, and we found no reputable outlet that gives the share.
  • Starship: said on 2026-04-28 that it had completed 10 million autonomous deliveries with more than 3,000 robots in 8 countries (press release), describing this as Level 4 without active human supervision; on 2026-06-04 it announced it is winding down U.S. university-campus operations and moving more than 1,200 campus robots to grocery retailers (press release). No remote-assistance share appears in its 2026 announcements.
  • Serve Robotics: its annual report for 2025 (10-K) gives a fleet of more than 2,000 robots at 2025-12-31 and describes them as Level 4 robots that can run without humans in the loop "for periods of time", backed by remote supervisors; daily active robots averaged 792 in Q2 2026 (8-K Exhibit 99.1).

These are all the companies' own statements. "Autonomous" means different things in different companies' material; before comparing, ask how many robots one remote operator watches, when they take over, and whether takeover counts are published. The same questions for cars are in the autonomous driving series.

What this series does not cover

Reading order

  1. Learning actions from demonstrations: mean squared error on the left-or-right obstacle, flow-matching sampling steps, the chunk-length trade-off, ALOHA, UMI, DROID, Open X-Embodiment, AgiBot World and first-person human video
  2. VLA models and sim-to-real: from text action tokens to FAST to flow-matching action experts, a hand-worked latency estimate, slow and fast layers, RL on VLAs, evaluation, sim-to-real, an openness table and humanoid status