Robots that learn by being shown: teleoperation, demonstrations and one-shot imitation. How many demonstrations a task needs, whether reinforcement learning must come on top, who supplies the rigs and who argues what, with a source for every fact.
Imitation learning, teaching a robot by showing it what to do, became the main way robots learn manipulation in the last three years. In 2023 ALOHA learned fine two-armed tasks from 50 demonstrations each, about ten minutes of data, at 80% to 90% success (arXiv 2304.13705); by 2026 companies claim new tasks from one short demonstration, and reliability above 99% on a single task such as folding laundry, all claims the companies make about their own models (Generalist; Sunday). Shared datasets grew from Open X-Embodiment's 22 robots (arXiv 2310.08864) to open models and a $100 arm for anyone (TechCrunch). The open questions are whether imitation alone is reliable enough or needs reinforcement learning on top, how many and which demonstrations a task needs, how much the robot inherits the operator's habits, whether it carries over to new sites and robot bodies, who corrects the robot once it is deployed, and whether demonstration data is a moat. How the data is captured (head cameras, gloves, hand-held grippers) is on the map's human data capture page.
Imitation alone reaches useful but not production rates on hard tasks: ACT 64% to 96% per task (arXiv 2304.13705); ALOHA Unleashed 40% to 75% after 26,000 demonstrations (arXiv 2410.13126); Gemini Robotics 2 at 36% to 89.6% on harder real tasks (company claim, Google DeepMind).
RL after imitation closes part of the gap: HIL-SERL near-perfect in 1 to 2.5 hours (arXiv 2410.21845); Recap doubles throughput and halves failures (company claim, arXiv 2511.14759).
But the highest published reliability figures come from imitation-style training on matched data, all company claims: Sunday 99.1% over 785 folds (Sunday), Dyna 99.4% over 24 hours (The Letter Two), Generalist 99% (The Robot Report).
Per-task counts fell: 50 for ACT and Mobile ALOHA (arXiv 2401.02117); 50 to 100 for Gemini Robotics On-Device (Google DeepMind); under 200 for a new robot with Gemini Robotics 2 (company claim); 176 examples (about 8 hours) for Physical Intelligence's sock task (our note); one to ten with phase decomposition (arXiv 2511.10110); one in context, at 59% to 79% on short tasks (arXiv 2606.08154; Generalist).
Diversity wins in three independent studies: environments and objects over demonstrations per setting (arXiv 2410.18647); task diversity over per-task quantity (arXiv 2507.06219); task diversity for in-context prompting (arXiv 2606.30457).
Volume still moves the base model: GEN-0's 270,000 hours and its 7B threshold (company claim); Figure's package time from 6.84 s to 4.31 s with 10 against 60 hours (company claim, Figure); Skild's in-context success rising with pretraining hours (company claim, Skild).
The robot inherits the operator: "What an operator can demonstrate bounds what imitation learning can learn"; motions awkward on an interface (exact orientations, multi-turn wrist rotations) are poorly demonstrated; triggerable primitives raised a real unscrewing task from 38% to 67% (arXiv 2512.04960).
Operator variation confounds learning: differences in speed and style between operators hurt, and correcting for them was worth 2.5 times the data (arXiv 2507.06219); novice teleoperators produce "task-successful but suboptimal" episodes with stalls, repeated corrections and joint-limit motions (arXiv 2605.26349).
Haptics: force feedback during demonstration lowered and steadied applied forces and made both demonstrations and the cloned policy faster, in a 10-person study (arXiv 2301.09157); force-feedback teleoperation took over 13 minutes to peel a zucchini that a hand-held force-capture tool did in 5, and force-aware data gave 54.5% relative improvement (arXiv 2410.07554).
New sites and objects: about 90% in new environments with enough environment diversity (arXiv 2410.18647); UMI 71.7% in unseen places (arXiv 2402.10329); π0.5 cleaning in unseen homes (arXiv 2504.16054); DROID +17% out of distribution (arXiv 2403.12945).
Benchmarks overstate it: LIBERO scores above 90% collapse under perturbation (arXiv 2510.03827; arXiv 2510.13626).
New robots: RT-X showed positive transfer across 22 robots (arXiv 2310.08864); Gemini Robotics 1.5 and 2 claim motion transfer and fast adaptation to new bodies (company claims); "Is Diversity All You Need" finds multi-robot pretraining optional, with single-robot data transferring (arXiv 2507.06219).
Corrections are now training data: HG-DAgger, Sirius, RLIF, HIL-SERL and RaC turn interventions into learning signal (arXiv 1810.02890; arXiv 2211.08416; arXiv 2311.12996; arXiv 2410.21845; arXiv 2509.07953); Recap uses corrections inside Physical Intelligence's training loop (arXiv 2511.14759).
Products ship with remote operators: 1X's NEO ($20,000 or $499 a month) launched with remote operators who can take over (Robotics 24/7, 28 Oct 2025).
Cost figures are vendor claims only: a teleoperation service vendor says early humanoid pilots "routinely sit at 30% intervention or above" and lists operators from $12 an hour (Adamo, 14 Aug 2026).
The open side grows: π0 weights (openpi), GR00T, Octo and OpenVLA are open; AIRoA released about 5,000 cleaned hours on 68 tasks with a π0.5-based model (AIRoA, 31 Jul 2026); AgiBot opened AGIBOT WORLD 2026 with error-recovery trajectories (The Robot Report, 7 Apr 2026).
Money also flows to proprietary data: Bessemer estimates industry robot data costs above $3 billion over two years and writes "Data is expensive, capital is the moat" (Bessemer, 16 Apr 2026); XDOF raised $70M for teleoperation data with 20 customers (TechCrunch, 17 Jun 2026); Scale AI reports more than 100,000 production hours of physical-AI data (company claim, Scale AI); Generalist raised $400M at $2B (SiliconANGLE, 4 Jun 2026).
Data ages with hardware: legacy data helps only above a competence threshold on new hardware (arXiv 2607.25593).
The Physical AI Voices by Herbert Scale Experts.