ZPE-Robotics

SAL v7.0 — free below $100M annual revenue. See LICENSE.

What This Is

187× compression on real robot joint streams (vs zstd_l19 4.59×, vs zstd_l3 4.44×) — with PrimitiveIndex search and VLA token export. Governing benchmark: LeRobot real-data suite, 3 datasets across 3 families. Proof path: proofs/enterprise_benchmark/GATE_VERDICTS.json.

ZPE-Robotics compresses and replays joint-stream logs for infrastructure teams that need searchable motion archives instead of raw bagfiles. The 187× figure is real and bounded-lossy (smooth-trajectory slices; step inputs cause Gibbs ringing — see footnote). Engineering surface remains blocker-governed: step-sequence behavior, comparator closure, and runtime closure are not closed.

Wire encoding: wire-v1 (frozen packet contract; authority: docs/ZPBOT_V2_AUTHORITY_SURFACE.md)

Key Metrics

Sources: proofs/enterprise_benchmark/GATE_VERDICTS.json (B2: 40.74× vs zstd_l19, B4/B5 latency) | proofs/red_team/red_team_report.json (42.14× vs zstd_l3, attack-5 FPR) | proofs/release_candidate/primitive_search_result.json (P@10) | proofs/release_candidate/anomaly_detection_result.json (FPR/recall) | proofs/release_candidate/it04_parity_matrix_result.json (cross-platform)

† Bounded-lossy. The ≤ 0.5° angular figure is limited to smooth-trajectory slices; it is not a general motion bound. Step/discontinuous inputs cause Gibbs ringing, with 68° RMSE measured on a unit-amplitude step signal. Baselines are lossless.

Competitive Benchmarks

Competitive benchmark evidence: proofs/enterprise_benchmark/benchmark_result.json | proofs/red_team/red_team_report.json | proofs/artifacts/lerobot_expanded_benchmarks/aggregate_spread_summary.json

† Bounded-lossy. The ≤ 0.5° angular figure is limited to smooth-trajectory slices; it is not a general motion bound. Step/discontinuous inputs cause Gibbs ringing, with 68° RMSE measured on a unit-amplitude step signal. All other baselines are lossless.

What We Prove

Auditable guarantees backed by committed proof artifacts. Start at docs/AUDITOR_PLAYBOOK.md.

• Spectral wire transport with directional reasoning layer for robot action sequences
• Search operates on decoded motion streams via PrimitiveIndex
• Red-team resilience: 4 attacks withstand, 1 fails, 1 partially withstands, 1 remains open
• VLA tokenization aligns with vision-language-action model input formats
• Public package acquisition route is pip install zpe-robotics; package availability does not change blocker status

What We Don't Claim

• Full release readiness
• Bit-exact .zpbot round-trip replay
• B3 benchmark gate pass
• Red-team attack 3 lossless qualification
• General anomaly readiness beyond the declared threshold-selected holdout surface
• Robotics Rust ABI
• Independent third-party reproduction
• Generally valid ≤ 0.5° angular fidelity — the figure comes from smooth-trajectory slices only; FFT-based encoding causes Gibbs ringing on step/discontinuous inputs (68° RMSE measured on a unit-amplitude step signal)
• Search-without-decode — PrimitiveIndex requires full packet decode before indexing

Commercial Readiness

Selected Claim Gate

Tests and Verification

Proof Anchors

• Runtime and proof artifacts outrank prose.
• GO and NO-GO language is reserved for named gates only.
• Historical proof bundles remain lineage only. They do not override the March 21 blocker-state evidence or the later Phase 10 threshold evidence.
• No IMC runtime import is introduced by this repo.
• The current installable package artifact is zpe-robotics 0.1.0, but the repo's blocker-state docs remain the authority surface for engineering status.

Use these files together:

Expanded LeRobot Benchmark Coverage

The expanded benchmark sweep is recorded in proofs/artifacts/lerobot_expanded_benchmarks/. The current qualified surface covers 3 real datasets across 3 materially distinct families, with compression ratios spanning 58.70x to 186.05x.

Read proofs/artifacts/lerobot_expanded_benchmarks/aggregate_spread_summary.json for the rollup verdict and proofs/artifacts/lerobot_expanded_benchmarks/dataset_manifest.json for the full attempt ledger, including qualification misses that were preserved rather than averaged away.

Repo Shape

The modality-lane count reflects the three recorded parity lanes (arm64-qemu, macos, ubuntu-x86) in proofs/release_candidate/it04_parity_matrix_result.json.

Quick Start

Install from PyPI:

pip install zpe-robotics
zpe-robotics --version

Or install from source (development):

pip install -e .
zpe-robotics --version

Repo-local engineering surface:

python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip setuptools wheel
python -m pip install -e ".[dev,benchmark,telemetry,netnew]"
python -m pytest tests -q
python -m build

If you need the shortest honest verification route, use docs/AUDITOR_PLAYBOOK.md. If you need the release workflow boundary, use proofs/runbooks/TECHNICAL_RELEASE_SURFACE.md.

Ecosystem

ZPE-Robotics follows the portfolio release-hygiene pattern used by ZPE-IMC, but it does not inherit IMC runtime claims, benchmark verdicts, or release readiness by association.

Observability: Comet dashboard (public)

Who This Is For

ZPE-Robotics is one of 17 independent encoding products in the Zer0pa portfolio — each domain-specific, each carrying its own proof surface. Robotics addresses motion telemetry; it does not inherit claims or release status from any other lane.