UR5e + Robotiq 2F-85
Six arm joints, eight articulated gripper coordinates and a free rigid block. Full source outcomes and matched controls are available.
Inspect the task replay ↗diff --git a/README.md b/README.md index 64a43a59fe2992b5961439faa9fed35a3895f1b6..412f3e9ac092a707416448fcb27447b16b7ac2df 100644 --- a/README.md +++ b/README.md @@ -43,7 +43,7 @@ denominator and every attempt remain under `replay/receipt/`. ```sh python3 -m venv .venv -.venv/bin/python -m pip install "dvidia-training[arm] @ git+https://github.com/Dvidia-Inference/dvidia-training.git@44fa7013d2718d8b28941bd82de37806f80ddf23" +.venv/bin/python -m pip install "dvidia-training[arm] @ git+https://github.com/Dvidia-Inference/dvidia-training.git@2a5fceaea8c2c7eeae808ea6656233d2a7f433fe" .venv/bin/dvidia-embodiment fetch-assets --output runs/reference-assets .venv/bin/dvidia-reference-task benchmark runs/reference-assets --output runs/grasp-development .venv/bin/dvidia-reference-task inspect runs/grasp-development @@ -56,7 +56,7 @@ physics and learning are separate workloads; no minimum RAM or speed guarantee has been established for this task. WebGL2 is required only for the interactive browser viewer. Fetch dependencies and pinned assets once, then native simulation and exported replay work offline through a local HTTP server. The complete -[local task guide](https://github.com/Dvidia-Inference/dvidia-training/blob/44fa7013d2718d8b28941bd82de37806f80ddf23/docs/native-task-benchmark.md) +[local task guide](https://github.com/Dvidia-Inference/dvidia-training/blob/2a5fceaea8c2c7eeae808ea6656233d2a7f433fe/docs/native-task-benchmark.md) describes controls, scoring and measured timing boundaries. ## Hardware profiles @@ -80,7 +80,7 @@ Frozen confirmation cohort: 9 planned attempts across 3 synthetic scenes. Counts - low_friction / replay_open_jaw: 0/1 completed, 1 numerically valid - low_friction / hold: 0/1 completed, 1 numerically valid -Authored controller, synthetic scenes and privileged native simulator state. Contact is uncalibrated; no learned skill or physical robot is qualified. This is separate from the unloaded PiPER mechanics bundle; no PiPER 3D viewer is included. Every attempted episode, including failed and invalid attempts, remains under `tasks/piper/attempts/`. Inspect [the complete task report](tasks/piper/report.html), [frozen protocol](tasks/piper/protocol.json), [file hashes](tasks/piper/integrity.json), and [original MIT model license](tasks/piper/licenses/agilex_piper/LICENSE). [Local PiPER contact guide](https://github.com/Dvidia-Inference/dvidia-training/blob/44fa7013d2718d8b28941bd82de37806f80ddf23/docs/piper-contact-pilot.md). +Authored controller, synthetic scenes and privileged native simulator state. Contact is uncalibrated; no learned skill or physical robot is qualified. This is separate from the unloaded PiPER mechanics bundle; no PiPER 3D viewer is included. Every attempted episode, including failed and invalid attempts, remains under `tasks/piper/attempts/`. Inspect [the complete task report](tasks/piper/report.html), [frozen protocol](tasks/piper/protocol.json), [file hashes](tasks/piper/integrity.json), and [original MIT model license](tasks/piper/licenses/agilex_piper/LICENSE). [Local PiPER contact guide](https://github.com/Dvidia-Inference/dvidia-training/blob/2a5fceaea8c2c7eeae808ea6656233d2a7f433fe/docs/piper-contact-pilot.md). ## PiPER guarded grip comparison @@ -111,8 +111,69 @@ Frozen confirmation cohort: 30 planned attempts across six synthetic scenes. Two Baseline and guarded authored controllers use privileged native state and their own matched open-jaw replays plus a common hold control. Full qualification decisions, unresolved outcomes, absent opportunities, alarms, retries and aborts remain in the source evidence. These synthetic trials are not detector accuracy, learned competence or physical qualification. -Every full attempt is stored as canonical deterministic gzip JSON under `tasks/piper-grip/attempts/`; the original seal records both compressed and decoded bytes and SHA-256. Download all files in `tasks/piper-grip/` together and run the pinned `dvidia-piper-grip inspect` command from [the local guide](https://github.com/Dvidia-Inference/dvidia-training/blob/44fa7013d2718d8b28941bd82de37806f80ddf23/docs/piper-grip-recovery.md). No archive extraction is needed. [Full comparison](tasks/piper-grip/report.html) · [Original protocol](tasks/piper-grip/protocol.json) · [Stored and decoded hashes](tasks/piper-grip/integrity.json). - +Every full attempt is stored as canonical deterministic gzip JSON under `tasks/piper-grip/attempts/`; the original seal records both compressed and decoded bytes and SHA-256. Download all files in `tasks/piper-grip/` together and run the pinned `dvidia-piper-grip inspect` command from [the local guide](https://github.com/Dvidia-Inference/dvidia-training/blob/2a5fceaea8c2c7eeae808ea6656233d2a7f433fe/docs/piper-grip-recovery.md). No archive extraction is needed. [Full comparison](tasks/piper-grip/report.html) · [Original protocol](tasks/piper-grip/protocol.json) · [Stored and decoded hashes](tasks/piper-grip/integrity.json). + + +## PiPER contact memory v0.3 + +Placement completion: baseline 4/12, probe-only 4/12, history 6/12. + +Retries, resumes and release refusals do not imply successful placement or recovery. + +Frozen confirmation cohort: 84 planned attempts across twelve synthetic scenes. Three frictions, two masses and two shared layouts; seven variants per scene. Every failed or invalid attempt remains in the denominator. Counts are descriptive, not physical reliability or detector accuracy. + +- μ=0.80, mass=0.04 kg / baseline_placement: 2/2 completed, 2 numerically valid +- μ=0.80, mass=0.04 kg / baseline_open_jaw: 0/2 completed, 2 numerically valid +- μ=0.80, mass=0.04 kg / probe_only_placement: 2/2 completed, 2 numerically valid +- μ=0.80, mass=0.04 kg / probe_only_open_jaw: 0/2 completed, 2 numerically valid +- μ=0.80, mass=0.04 kg / history_placement: 2/2 completed, 2 numerically valid +- μ=0.80, mass=0.04 kg / history_open_jaw: 0/2 completed, 2 numerically valid +- μ=0.80, mass=0.04 kg / hold: 0/2 completed, 2 numerically valid +- μ=0.80, mass=0.06 kg / baseline_placement: 0/2 completed, 2 numerically valid +- μ=0.80, mass=0.06 kg / baseline_open_jaw: 0/2 completed, 2 numerically valid +- μ=0.80, mass=0.06 kg / probe_only_placement: 0/2 completed, 2 numerically valid +- μ=0.80, mass=0.06 kg / probe_only_open_jaw: 0/2 completed, 2 numerically valid +- μ=0.80, mass=0.06 kg / history_placement: 2/2 completed, 2 numerically valid +- μ=0.80, mass=0.06 kg / history_open_jaw: 0/2 completed, 2 numerically valid +- μ=0.80, mass=0.06 kg / hold: 0/2 completed, 2 numerically valid +- μ=0.55, mass=0.04 kg / baseline_placement: 2/2 completed, 2 numerically valid +- μ=0.55, mass=0.04 kg / baseline_open_jaw: 0/2 completed, 2 numerically valid +- μ=0.55, mass=0.04 kg / probe_only_placement: 2/2 completed, 2 numerically valid +- μ=0.55, mass=0.04 kg / probe_only_open_jaw: 0/2 completed, 2 numerically valid +- μ=0.55, mass=0.04 kg / history_placement: 2/2 completed, 2 numerically valid +- μ=0.55, mass=0.04 kg / history_open_jaw: 0/2 completed, 2 numerically valid +- μ=0.55, mass=0.04 kg / hold: 0/2 completed, 2 numerically valid +- μ=0.55, mass=0.06 kg / baseline_placement: 0/2 completed, 2 numerically valid +- μ=0.55, mass=0.06 kg / baseline_open_jaw: 0/2 completed, 2 numerically valid +- μ=0.55, mass=0.06 kg / probe_only_placement: 0/2 completed, 2 numerically valid +- μ=0.55, mass=0.06 kg / probe_only_open_jaw: 0/2 completed, 2 numerically valid +- μ=0.55, mass=0.06 kg / history_placement: 0/2 completed, 2 numerically valid +- μ=0.55, mass=0.06 kg / history_open_jaw: 0/2 completed, 2 numerically valid +- μ=0.55, mass=0.06 kg / hold: 0/2 completed, 2 numerically valid +- μ=0.15, mass=0.04 kg / baseline_placement: 0/2 completed, 2 numerically valid +- μ=0.15, mass=0.04 kg / baseline_open_jaw: 0/2 completed, 2 numerically valid +- μ=0.15, mass=0.04 kg / probe_only_placement: 0/2 completed, 2 numerically valid +- μ=0.15, mass=0.04 kg / probe_only_open_jaw: 0/2 completed, 2 numerically valid +- μ=0.15, mass=0.04 kg / history_placement: 0/2 completed, 2 numerically valid +- μ=0.15, mass=0.04 kg / history_open_jaw: 0/2 completed, 2 numerically valid +- μ=0.15, mass=0.04 kg / hold: 0/2 completed, 2 numerically valid +- μ=0.15, mass=0.06 kg / baseline_placement: 0/2 completed, 2 numerically valid +- μ=0.15, mass=0.06 kg / baseline_open_jaw: 0/2 completed, 2 numerically valid +- μ=0.15, mass=0.06 kg / probe_only_placement: 0/2 completed, 2 numerically valid +- μ=0.15, mass=0.06 kg / probe_only_open_jaw: 0/2 completed, 2 numerically valid +- μ=0.15, mass=0.06 kg / history_placement: 0/2 completed, 2 numerically valid +- μ=0.15, mass=0.06 kg / history_open_jaw: 0/2 completed, 2 numerically valid +- μ=0.15, mass=0.06 kg / hold: 0/2 completed, 2 numerically valid + +Baseline, probe-only and history are authored controllers with privileged native state, each with its own matched open-jaw replay and a common hold control. Full diagnostics, contact-memory decisions, refusals, retries, resumes and aborts are retained. These synthetic trials establish no learned competence, physical reliability or detector accuracy. + +Every attempt is canonical deterministic gzip JSON under `tasks/piper-contact-v03/attempts/`; the original seal records both stored and decoded bytes and SHA-256. Download all files in `tasks/piper-contact-v03/` together for the pinned offline inspection workflow in [the v0.3 guide](https://github.com/Dvidia-Inference/dvidia-training/blob/2a5fceaea8c2c7eeae808ea6656233d2a7f433fe/docs/piper-contact-memory-v03.md). [Full comparison](tasks/piper-contact-v03/report.html) · [Every outcome](tasks/piper-contact-v03/report.json) · [Original protocol](tasks/piper-contact-v03/protocol.json) · [Stored and decoded hashes](tasks/piper-contact-v03/integrity.json). + + +The reviewed publication now permits at most **256 MiB of stored files**, so the v0.3 +comparison can retain all previous evidence alongside its 84 new full attempts. +The **32 MiB per-file limit** and bounded single-member gzip decoding remain in place. +This is a packaging limit, not a hardware or training memory requirement. ## Provenance and licensing diff --git a/index.html b/index.html index 0a28cb6ca7c8c7d289a87ae047af9921179040a9..0179c384ee0d595e84302d69e40a98fac8831a59 100644 --- a/index.html +++ b/index.html @@ -5,6 +5,6 @@
A native six-joint arm, an articulated gripper and a free rigid object. Recorded task attempts make the simulator's behavior visible, and the complete receipts make its limits checkable.
This is recorded simulation from an authored controller with privileged simulator state. The browser does not run physics or connect to hardware. Learned skills, shoe manipulation and physical transfer remain future stages.
Use orbit, zoom, recorded motion selection and the timeline inside the viewer. Replay is an excerpt; the full benchmark includes every matched control and invalid attempt. The native meshes, binary poses, complete receipts and license notices are served locally by this Space.
Six arm joints, eight articulated gripper coordinates and a free rigid block. Full source outcomes and matched controls are available.
Inspect the task replay ↗Six arm commands in radians and one jaw command in metres. This pilot uses an authored controller and privileged native simulator state. Contact is uncalibrated; no learned skill, shoe competence or physical robot is qualified. No PiPER 3D viewer is included.
Placement completion: baseline 4/6 versus guarded 2/6. This cohort does not show improved placement success.
Retries and qualification decisions do not imply successful recovery.
Frozen confirmation cohort: 30 planned attempts across six synthetic scenes. Two shared layouts at each friction; five variants per scene. Counts are descriptive, not physical reliability or detector accuracy.
μ=0.80 / baseline_placement: 2/2 completed, 2 numerically validBaseline and guarded controllers use privileged native state. Each has its own matched open-jaw replay; a shared hold control remains. Qualification decisions, retries, alarms, unresolved opportunities and aborts are retained. These are authored policies with uncalibrated contact, not learned skills or a physically qualified detector.
Every full attempt is retained as deterministic gzip JSON. The CLI checks compressed and decoded hashes with bounded decompression; no archive extraction is needed. This comparison adds numerical evidence to the existing contact pilot; it adds no PiPER 3D viewer.
6 arm axes + 1 gripper axis. This separate bundle measures native model structure and authored unloaded motion. Its numerical receipts do not qualify a contact task, learned policy or physical robot. Simulation asset license: MIT.
Pinned local simulator guide ↗5 arm axes + 1 gripper axis. This separate bundle measures native model structure and authored unloaded motion. Its numerical receipts do not qualify a contact task, learned policy or physical robot. Simulation asset license: Apache-2.0.
Pinned local simulator guide ↗The roadmap extends to bilateral manipulation and five-finger hands. Contact tasks, sensing adapters and physical qualification remain to be built.
Follow the research ↗A native mechanics benchmark establishes model structure and recorded motion. A native task benchmark adds measured contact outcomes. Neither establishes hardware reliability or a learned motor policy. Upcoming profiles have no verified implementation claim.
The native task is a local CPU workload. Fetch the pinned dependencies and assets once, then simulate offline. WebGL2 supports viewing; no hosted inference or paid GPU server is requested by this static Space. Minimum RAM and arbitrary-dataset speed guarantees remain unmeasured.
Six arm joints, eight articulated gripper coordinates and a free rigid block. Full source outcomes and matched controls are available.
Inspect the task replay ↗Placement completion: baseline 4/12, probe-only 4/12, history 6/12.
Retries, resumes and release refusals do not imply successful placement or recovery.
Frozen confirmation cohort: 84 planned attempts across twelve synthetic scenes. Three frictions, two masses and two shared layouts; seven variants per scene. Every failed or invalid attempt remains in the denominator. Counts are descriptive, not physical reliability or detector accuracy.
Three authored controllers use privileged native simulator state. Each placement policy has its own matched open-jaw replay; a shared hold control remains. Full contact history, decisions, refusals, retries, resumes, aborts and invalid outcomes are retained. Uncalibrated contact proxies do not qualify a learned skill, physical detector or robot.
All eighty-four full attempts are retained as deterministic gzip JSON with original stored and decoded hashes. Previous task and numerical evidence remains separate and unchanged. This adds no PiPER 3D viewer or physical qualification.
Six arm commands in radians and one jaw command in metres. This pilot uses an authored controller and privileged native simulator state. Contact is uncalibrated; no learned skill, shoe competence or physical robot is qualified. No PiPER 3D viewer is included.
Placement completion: baseline 4/6 versus guarded 2/6. This cohort does not show improved placement success.
Retries and qualification decisions do not imply successful recovery.
Frozen confirmation cohort: 30 planned attempts across six synthetic scenes. Two shared layouts at each friction; five variants per scene. Counts are descriptive, not physical reliability or detector accuracy.
μ=0.80 / baseline_placement: 2/2 completed, 2 numerically validBaseline and guarded controllers use privileged native state. Each has its own matched open-jaw replay; a shared hold control remains. Qualification decisions, retries, alarms, unresolved opportunities and aborts are retained. These are authored policies with uncalibrated contact, not learned skills or a physically qualified detector.
Every full attempt is retained as deterministic gzip JSON. The CLI checks compressed and decoded hashes with bounded decompression; no archive extraction is needed. This comparison adds numerical evidence to the existing contact pilot; it adds no PiPER 3D viewer.
6 arm axes + 1 gripper axis. This separate bundle measures native model structure and authored unloaded motion. Its numerical receipts do not qualify a contact task, learned policy or physical robot. Simulation asset license: MIT.
Pinned local simulator guide ↗5 arm axes + 1 gripper axis. This separate bundle measures native model structure and authored unloaded motion. Its numerical receipts do not qualify a contact task, learned policy or physical robot. Simulation asset license: Apache-2.0.
Pinned local simulator guide ↗The roadmap extends to bilateral manipulation and five-finger hands. Contact tasks, sensing adapters and physical qualification remain to be built.
Follow the research ↗A native mechanics benchmark establishes model structure and recorded motion. A native task benchmark adds measured contact outcomes. Neither establishes hardware reliability or a learned motor policy. Upcoming profiles have no verified implementation claim.
The native task is a local CPU workload. Fetch the pinned dependencies and assets once, then simulate offline. WebGL2 supports viewing; no hosted inference or paid GPU server is requested by this static Space. Minimum RAM and arbitrary-dataset speed guarantees remain unmeasured.