Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 12 additions & 0 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -26,6 +26,18 @@ All four covers are **AI-generated editorial illustrations**, not footage, measu

## Latest studies

[PiPER contact recovery v0.4](reports/PiPER%20contact%20recovery%20v04.md)
traces all six earlier failures to acquisition under the pinned position servo
and tests carry recovery separately from avoiding an unchanged supported retry.
Fresh confirmation retains **84 valid attempts**, with **4/12 completed placements in
every mode and 0/48 negative-control successes**. Two resumed placements finish
2.440–2.640 simulated seconds sooner while keeping their original settling holds.
Four failed acquisitions reach refusal 2.88–2.90 seconds sooner; no completion is
gained or lost. Fresh layouts expose two held-but-too-low resume failures in every
mode, alongside six acquisition failures. The [complete receipt](topics/physical-grounding/runs/2026-10-11-piper-recovery-v04/README.md)
retains every attempt and all 84,000 raw computation durations; hardware remains
unqualified.

[PiPER contact history v0.3](reports/PiPER%20contact%20history%20v03.md)
compares original motion, an identical acquisition probe without carry monitoring,
and causal contact history with requested slowdown, closed reassessment and resume.
Expand Down
2 changes: 2 additions & 0 deletions docs/downloads/reports/PiPER contact history v03.md
Original file line number Diff line number Diff line change
Expand Up @@ -128,3 +128,5 @@ writing follows this repository's publication license.

*Published by DVIDIA. Contact [hello@dvidia.org](mailto:hello@dvidia.org) and follow
[@iammrriver](https://x.com/iammrriver).*

The follow-up [contact recovery v0.4 study](https://research.dvidia.org/papers/piper-contact-recovery-v04/) diagnoses these six acquisition failures and compares carry changes separately from the decision to repeat a supported grasp. The complete original v0.3 evidence above remains unchanged.
113 changes: 113 additions & 0 deletions docs/downloads/reports/PiPER contact recovery v04.md

Large diffs are not rendered by default.

Original file line number Diff line number Diff line change
@@ -0,0 +1,80 @@
# PiPER contact recovery v0.4 — frozen confirmation

Fresh seed **941827**, runtime [`ccf8f3d1c2a1977da855ba8f601499aa543a7b6c`](https://github.com/Dvidia-Inference/dvidia-training/tree/ccf8f3d1c2a1977da855ba8f601499aa543a7b6c).

The fresh 84-attempt confirmation completes **4/12 placements in each mode**:
frozen v0.3 history, v0.4 carry-only and v0.4 full. All attempts are numerically
valid and all 48 negative controls fail. The two completed placements that
resume after a hold finish **2.440 and 2.640 simulated seconds sooner** in v0.4;
two other completions are unchanged. Four low-friction failures are refused
**2.88–2.90 seconds sooner**, while remaining failed placements. There are no
gained or lost completions in the matched comparison.

| Friction | Mass | v0.3 history | v0.4 carry-only | v0.4 full | Matched open-jaw + hold controls |
| --- | ---: | ---: | ---: | ---: | ---: |
|0.8|40 g|2/2|2/2|2/2|0/8|
|0.8|60 g|0/2|0/2|0/2|0/8|
|0.55|40 g|2/2|2/2|2/2|0/8|
|0.55|60 g|0/2|0/2|0/2|0/8|
|0.15|40 g|0/2|0/2|0/2|0/8|
|0.15|60 g|0/2|0/2|0/2|0/8|
|Total|—|4/12|4/12|4/12|0/48|

The new nominal-friction 60 g failures regain bilateral contact during the hold,
but the object settles below the unchanged 6 mm resume floor before sustaining
100 ms of eligible dwell. Both remain suspended and time out without resuming.
Their full native states and commands are equal across all three modes. These
fresh failures expose layout sensitivity; the earlier 6/12 v0.3 result is a
different cohort and is not reused as a confirmation result here.

The two successful resumed tasks keep the same 220 and 240 ms holds; the improvement
comes after settling, from restoring requested movement. The mean paired time
change over all four shared successes is −1.270 simulated seconds. Four low-friction
refusals skip the unchanged retry, but do not become successful placements. The
moderate-friction 60 g cases retain the original unsuccessful retry.

## Complete original evidence

`receipt/` is the complete, byte-identical 90-file benchmark bundle: 84 compressed
attempts, protocol, report JSON/HTML, original seal and MIT model notices.
All 84 attempts retain 20 simulated seconds, 1,000 commands and 1,001 native states.
Every raw per-decision computation time is included; 0/84,000 recorded windows
exceed 20 ms on this host. All warnings and prohibited-contact counts are zero.

- [Readable paired report](receipt/report.html)
- [Every outcome and paired timing](receipt/report.json)
- [Frozen protocol](receipt/protocol.json)
- [Stored and decoded hashes](receipt/integrity.json)
- [Execution boundary](execution.json)
- [Installed runtime source check](installed-runtime.json)
- [Publication provenance](provenance.json)
- [Verified Hugging Face publication](huggingface-publication.json)

Seal-listed files total 58,450,596 stored bytes and 327,061,224 bytes with attempts
decoded. Stored report SHA-256 is
`899a05b44e9b76a597fb2683d433e445d84ae485ed95d15de463f3dab3173164`;
original seal SHA-256 is
`b0f6c0302fce5a44f4afe4bbe2451875b1c7a049340f08911465171958ec5156`.
Execution, installed-runtime and publication records are outside the original
bundle seal. Their presence does not rewrite or selectively shorten any attempt.

The installed source archive runs with OS process-level network denial and local
pinned assets. Whole-command wall time is 133.629 seconds for 1,680 simulated seconds;
it includes evaluation, compression and inspection. Attempt wall times are
1.175–1.728 seconds, per-attempt control p95 is 1.087–3.181 ms, and the worst raw
control window is 13.473 ms. The host is macOS 26.5.2 arm64, 10 logical CPUs,
Python 3.12.9, MuJoCo 3.15.0 and NumPy 2.5.3; no GPU is used, and OpenBLAS/OpenMP each
use one thread. Process high-water RSS is 727.19 MiB before inspection; the child
lifetime including inspection reaches 738.41 MiB. Host load is uncontrolled.

These are software computation times and measured memory peaks, not physical
sensor-to-actuator latency or universal minimum hardware requirements. Simulation
contacts, geometry and material values remain uncalibrated; this is an authored
privileged-state controller, not a learned Skillspace skill or hardware qualification.
Numerical validity does not imply successful placement. Original DVIDIA code and
the pinned PiPER model retain MIT terms.

[Original six-failure diagnosis](forensics/README.md) · [Source and development archive](https://github.com/Dvidia-Inference/dvidia-training/tree/ccf8f3d1c2a1977da855ba8f601499aa543a7b6c/benchmarks/piper-recovery-development-20261011) · [Robot Simulation Lab](https://huggingface.co/spaces/Dvidia/robot-simulation-lab)


The full hosted bundle is verified at [Hugging Face commit be267fe5a2419e2b0d8098b5126f5f351d7df48f](https://huggingface.co/spaces/Dvidia/robot-simulation-lab/tree/be267fe5a2419e2b0d8098b5126f5f351d7df48f/tasks/piper-recovery-v04). All 293 payload files were checked at that immutable revision. Earlier evidence remains unchanged.
Original file line number Diff line number Diff line change
@@ -0,0 +1,16 @@
{
"schema": "dvidia.local-benchmark-execution/v1",
"source_commit": "ccf8f3d1c2a1977da855ba8f601499aa543a7b6c",
"cohort": "confirmation",
"seed": 941827,
"returncode": 0,
"execution": "Installed wheel built from a committed source archive; runtime source paths clean; OS process-level network denial; local pinned assets; PYTHONPATH unset",
"threads": {
"OPENBLAS_NUM_THREADS": "1",
"OMP_NUM_THREADS": "1"
},
"command_wall_seconds": 133.62855662498623,
"child_lifetime_peak_rss_bytes": 774275072,
"memory_scope": "macOS RUSAGE_CHILDREN high-water RSS including benchmark and inspection; preceding Git children also belong to child lifetime",
"wall_scope": "Whole installed command including startup, 84 native attempts, evaluation, compression, serialization and final inspection. Excludes wheel installation, asset provisioning, browser rendering and publication. Host load uncontrolled."
}
Original file line number Diff line number Diff line change
@@ -0,0 +1,35 @@
# Why six v0.3 PiPER placements failed

All six failures occurred during acquisition, before the contact-history carry controller had anything to recover. All alarm snapshots were table-supported; 11 of 12 probe windows remained supported throughout, with one unsupported snapshot during the remaining retry. The object moved at most **0.040–0.148 mm** across each complete episode while the first probe raised the TCP **3.428–3.849 mm**. Bilateral contact occupied 94.7–100% of each probe window; occasional single-sample contact gaps did not lead to acquisition. The retry repeated the same grip and failed again.

| Scene | First alarm (s) | Object / TCP rise at alarm (mm) | Final abort (s) | Retry overhead (s) | Episode max rise (mm) |
|---|---:|---:|---:|---:|---:|
| friction015-mass040-layout00 | 11.52 | 0.050 / 3.428 | 14.42 | 2.90 | 0.080 |
| friction015-mass040-layout01 | 11.50 | 0.052 / 3.448 | 14.40 | 2.90 | 0.080 |
| friction015-mass060-layout00 | 11.54 | 0.018 / 3.849 | 14.44 | 2.90 | 0.044 |
| friction015-mass060-layout01 | 11.50 | 0.016 / 3.435 | 14.40 | 2.90 | 0.040 |
| friction055-mass060-layout00 | 11.56 | 0.104 / 3.644 | 14.48 | 2.92 | 0.148 |
| friction055-mass060-layout01 | 11.54 | 0.088 / 3.680 | 14.46 | 2.92 | 0.142 |

The first alarm is `object_not_following` in all six cases. The second alarm is the same in five; μ=.55, 60 g, layout01 reaches `relative_slip` first. These names indicate which condition won the guard’s priority ordering, not independent ground-truth slip labels. In that last case the object is still table-supported and rises only 0.133 mm.

## Why more waiting or the same retry cannot add grip force

The upstream gripper is a position servo: `F = 40(target − q) − 5 q̇`, before its ±10 N clamp. At a stationary finger slide around 0.0196 m, the legal minimum target is already zero, so the drive supplies about −0.784 N. The coupled pair carries roughly 0.39 N per side. The actuator force cap is **not** being hit; it is the unchanged position target / gain / grasp combination that establishes this squeeze.

For approximately vertical object sides, a useful static support estimate is `μ × (Nleft + Nright)`. At μ=.15 this is about 0.118 N against weights 0.392 N (40 g) and 0.589 N (60 g). At μ=.55 it is about 0.43 N against 0.589 N for the failed 60 g cases. The .55 / 40 g successes have just enough approximate static support; .8 / 60 g also qualifies but needs carry recovery.

The compiled native contact check confirms both tangential friction coefficients equal the scene value; the object’s priority1 supplies them. First-alarm contact normals have |z| ≤0.00204, consistent with nearly vertical side contacts. This check evaluates parameters and geometry at recorded poses only. It does not substitute reconstructed normal forces for recorded values.

This is a diagnostic approximation, not a proof that every alternate grasp is impossible. It omits sub-threshold logged forces and does not reconstruct tangential wrenches. Hardware squeeze control, reorientation, support-assisted manipulation or a different grasp could alter feasibility and would need their own declared interface and benchmark.

## What v0.4 can honestly improve

1. Recognize a supported failed probe with persistently loaded pads and the jaw already closed, then stop repeating an unchanged grasp. The old retry adds **2.90–2.92 simulated seconds** before reaching the same refusal. The controller should use only observed history, not hidden mass or friction.
2. Keep the acquisition and task thresholds unchanged. A quicker refusal is a useful diagnostic result, but remains a failed placement.
3. Improve carry recovery separately, where the six successful cases demonstrate that acquisition is feasible. Remove unnecessary time under the risk condition, measure actual velocity, and avoid making “smaller requested joint delta” synonymous with slower motion.
4. Preserve false-positive and regression opportunities: same supported condition could be temporary contact settling. Test transient disturbances and ambiguous windows; retain retry for conditions where a new grasp could plausibly change contact.

All six failed history trajectories match their probe-only counterpart in every recorded qpos, qvel and command. All six unguarded baselines also fail, with maximum lift only 0.108–0.301 mm. This corroborates an acquisition limitation rather than a v0.3 carry-history regression.

`analysis.json` contains each event, both probe windows, normal-force/jaw statistics, source checksums, and the native contact-parameter check. `analyze.py` reproduces this analysis from the frozen evidence. This known v0.3 cohort is development data for v0.4; fresh confirmation must be separate.
Loading
Loading