Measured product lead custody receipt.
github.com/Zer0pa/ZPE-RoboticsDEVELOPER-READY
LIVE
B3 OPEN
Robots That Learn Moves Like Humans Do.
A movement memory for robots — the form of an action, kept · ZPE-Robotics · PyPI zpe-robotics v0.1.1 · github.com/Zer0pa/ZPE-Robotics
A person learns a waltz, a kung fu form, or how to pick something up the same way — by repeating the movement until its shape settles into the body.
What stays is not one attempt; it is the form. Robots have never had a memory for that. ZPE-Robotics is one: it keeps the form of an action — pick, wipe, push, pull — so a robot can hold a movement, search it, and learn from it. Proven on smooth motion, real LeRobot data, at 187.13×.
01 · THE GAPRECORDED, NOT LEARNED
A robot records a movement perfectly, yet cannot learn it. a recording is not a memory.
02 · MARKETSADJACENT FORECASTS
Robot software — ’31 — $67.9B · Digital twin — ’30 — $155.8B · Warehouse robotics — ’30 — $17.3B · Industrial robotics — ’30 — $16.5B · AMR — ’30 — $8.7B · source: Next Move Strategy, MarketsandMarkets
Every robot that learns to move works inside these markets; ZPE-Robotics is the memory beneath them.
03 · VALUE
187.13×
Compression vs zstd_l19 on real LeRobot joint streams · bounded-lossy smooth motion
04 · INSIGHT
Practiced enough, a movement leaves one thing behind: its form.
05.1 · CURRENT TECHRECORDED AND SHELVED
Today a robot's movement gets dumped into ROS bagfiles or parquet. The files are large, findable only by timestamp or filename, never by the movement itself. Nothing downstream can learn from a recording it cannot read.
05.2 · OUR TECHKEEP THE FORM
ZPE-Robotics keeps the form. It encodes a robot's movement into a bounded-lossy archive — keeping the shape of the action, dropping the once-only noise — at 187.13× on real LeRobot data. PrimitiveIndex returns runs by the movement inside them: every clean reach, every dropped grasp, every recovered pour. Pick, wipe, push, pull become findable, not just stored.
05.3 · BENCHMARKSLEROBOT REAL DATA
Compression187.13× vs zstd_l19
Encode P500.111ms / 1k frames
Decode P500.089ms / 1k frames
Checks4/5archive suite
Scope: 3 LeRobot datasets; 58.70–187.13× spread. General replay and search remain open.
06 · MEASUREMENTMEASURED ARCHIVE SURFACE
Archive claims stay tied to real LeRobot slices and smooth-motion limits.
06.1 · COMPARATIVE PERFORMANCE · LEROBOT BYTES PER FRAME
LeRobot declared episodes (columbia_cairlab_pusht_real, 136 episodes, 27,808 frames), smooth-trajectory slices. Baselines are lossless zstd, gzip, lz4, MCAP, HDF5 variants. Spread across 3 datasets: 58.70–187.13×, median 61.27×. Source: proofs/enterprise_benchmark/benchmark_result.json.
07 · KEY METRICSMEASURED RESULTS
07.1 · COMPRESSION
187.13×
vs zstd_l19 4.59× · bounded-lossy LeRobot data
07.2 · ENCODE P50
0.111ms
per 1k frames · check B4 PASS
07.3 · DECODE P50
0.089ms
per 1k frames · check B5 PASS
07.4 · ARCHIVE CHECKS
4 / 5PASS
smooth archive PASS · general replay open
07.5 · DATASET SPREAD
61.27×
median of 3 LeRobot datasets · 187.13× peak
08 · REPLAY FIDELITYSMOOTH VS STEP
Smooth movement stays inside the archive boundary; stepped movement does not.
08.1 · WHAT THE ARCHIVE SUPPORTSSMOOTH SLICE
On smooth-trajectory slices of declared LeRobot data, movement encodes and decodes consistently across arm64, macOS and x86. A sharp or stepped movement does not: the FFT-based encoder rings — Gibbs distortion — measured at 68° RMSE on a unit-amplitude step. A step has no smooth form to keep.
Search-without-decode and general bit-level replay remain open. PrimitiveIndex still walks decoded streams. The credibility claim is bounded-lossy smooth movement — useful for archive, analysis, and downstream teaching, not for live closed-loop control where every byte of the motion has to come back exactly.
08.2 · HONEST BLOCKER
Honest Blocker ·
187.13× is bounded-lossy on smooth movement; sharp, stepped movement still rings. General replay and search-without-decode are false. PrimitiveIndex requires decode. PyPI v0.1.1 is stale; zpe-motion-kernel is legacy; no Robotics Rust ABI. RT3 miss, RT4 partial, RT7 open.
09
Movement becomes memory inside the archive.
09.1 · THE AMBITION
The aim is not a better robot policy — it is the memory underneath one. A robot that keeps the form of a movement can recall it, refine it, and pass it on. Demonstration stops being disposable capture and starts behaving like inventory a fleet can build on.
09.2 · WHAT WORKS NOW
On smooth movement: 187.13× archives and recall follow the action's shape.
09.3 · WHAT'S STILL OPEN
Still open: bit-level replay, sharp-movement distortion, search without decode, independent reproduction, a current release.
09.4 · REPERTOIRE · NEAR-TERM (12–24 MO)
A robot keeps every taught movement
A teleoperation team that used to throw away demonstrations after training can now keep every pick, wipe, push, and pull. At 187.13× on smooth motion, a humanoid's entire taught repertoire fits in the space its raw logs used to take for one afternoon.
09.5 · RECALL · NEAR-TERM (12–24 MO)
Engineers find runs using movement shape
A robotics platform engineer hunting a specific failure mode stops scrubbing video and grepping bag files. The archive returns every clean reach, every dropped grasp, every retry by the shape of the action — so the question “show me the bad pours” gets a direct answer.
09.6 · TEACHING · MID-TERM (24–48 MO)
One robot's motion teaches the next
A humanoid R&D lead exporting movements as vision-language-action tokens hands a taught skill straight into the next model generation. The form one robot kept after a thousand pours becomes the starting condition for the robot that hasn't poured anything yet.
09.7 · SIMULATION · MID-TERM (24–48 MO)
Simulation gets back real demonstrations
Once replay closes for stepped motion, a simulation team can rerun the actual factory floor inside their environment — the dropped boxes, the missed grasps, the recoveries — instead of synthesising plausible ones. Sim and reality converge around the same retained movement.
09.8 · APPRENTICESHIP · PARADIGM (48 MO+)
Robots learn the way apprentices do
When movement can be kept, searched, and faithfully replayed, a robot stops being trained by exposure and starts being taught the way a person learns a craft — holding each form, refining it across attempts, passing it to the next robot the way a master hands down a technique.