Robotics research · Image goals / 0.1
Does it only look finished? A goal image describes the destination. This experiment asks Jev to judge the current scene, then compares its answer with the simulator’s physical state. Missing shoes and misleading arrangements stay in the results.
Loading the recorded experiment.
Enable JavaScript to inspect individual cases. The complete public report remains available as JSON below.
Inspect every case Show All cases Missing target Physically incomplete Physically complete False completion Request errors Not run
Current observation Goal reference
Physical reference Recorded simulator labels. A single image may not reveal hidden contact or sustained support.
Jev judgment A remote research judgment. This page sends no images, runs no model and commands no robot.
Count the misses. Keep the clock honest. Judgment accuracy uses the frozen, observability-aware verdict for every planned case: complete, incomplete, or unknown. A correct abstention matters when an image cannot establish hidden motion, contact or dwell. Errors and cases not run remain in the planned denominator; no score is claimed before requests are attempted. Physical false-completion and missing-target counts retain every planned negative case, with answered coverage shown separately.
Decision check Correct Denominator
Latency scope Timed requests p50 p95 p99 Maximum
Timing is recorded wall-clock latency, not the robot’s reaction time or simulator timestep. Quantiles use linear interpolation and retain timed errors in the all-attempt row. A small cohort cannot establish production tail latency. Missing measurements remain unreported.
This benchmark uses synthetic simulation fixtures and a hosted decision model. It does not publish trained robot weights or demonstrate hardware transfer. The DVIDIA code license does not grant rights to the DVIDIA marks or to an external model.
See the motion. Inspect the evidence. 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.
Frozen confirmation cohort: 18 planned attempts across 6 synthetic scenes. Counts are descriptive, with no population or physical reliability claim.
-A live PiPER shoe workstation. Drag the gripper target, open and close the jaws, and push a free rigid shoe. MuJoCo physics runs locally in your browser. Explore contacts and download a session receipt.
Authored control and synthetic shoe properties; no learned policy or hardware qualification.
Enter the live workstation → The shoe asset library is runnable. Choose one of three robot profiles, inspect six original rigid shoe proxies, download a reusable scene and run CPU simulation locally. Synthetic properties are labelled as assumptions; gravity/support benchmarks retain every attempt and matched control.
Open the asset library → Does it only look finished? Explore 32 frozen PiPER image-goal cases, including absent targets, wrong arrangements and visually identical moving scenes. Native fixture validation is published. Jev accuracy and service latency are not yet measured. Synthetic diagnostic fixtures; passive evaluation; no robot controller or hardware qualification.
Inspect image goals and physical references → Give the gripper a fair test. Compare a small block, a simplified shoe and the current shoe across light and heavy loads, with the rake mounted or absent. One shared authored policy and independent retention checks expose how the setup affects the lift. Recorded native physics; rigid synthetic objects; no physical robot qualification.
Explore the graspability comparison → From reaching a shoe to keeping hold. Compare a fixed grasp sequence with continuous acquisition checks, a measured lift checkpoint and bounded grip-loss recovery. Recorded native MuJoCo attempts preserve the full comparison, failures and independent scoring. Authored policies; synthetic contact; no physical robot qualification.
Inspect the grip comparison and every result → Before a skill can succeed. Explore a hinged shoe box and an out-of-reach shoe. Recorded native MuJoCo attempts retain failed access, withdrawal and tool-assisted reach. Authored policies; synthetic properties; no physical robot qualification.
Inspect both simulations and every result → Native grasp, carry and release.
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.
A live PiPER shoe workstation. Drag the gripper target, open and close the jaws, and push a free rigid shoe. MuJoCo physics runs locally in your browser. Explore contacts and download a session receipt.
Authored control and synthetic shoe properties; no learned policy or hardware qualification.
Enter the live workstation → The shoe asset library is runnable. Choose one of three robot profiles, inspect six original rigid shoe proxies, download a reusable scene and run CPU simulation locally. Synthetic properties are labelled as assumptions; gravity/support benchmarks retain every attempt and matched control.
Open the asset library → How does Jev judge completion? Explore 32 frozen PiPER image-goal cases, including absent targets, wrong arrangements and visually identical moving scenes. Inspect all paired Jev judgments, failures, abstentions and measured service latency. Synthetic diagnostic fixtures; passive evaluation; no robot controller or hardware qualification.
Inspect image goals and physical references → Does it only look finished? Explore 32 frozen PiPER image-goal cases, including absent targets, wrong arrangements and visually identical moving scenes. Native fixture validation is published. Jev accuracy and service latency are not yet measured. Synthetic diagnostic fixtures; passive evaluation; no robot controller or hardware qualification.
Inspect image goals and physical references → Give the gripper a fair test. Compare a small block, a simplified shoe and the current shoe across light and heavy loads, with the rake mounted or absent. One shared authored policy and independent retention checks expose how the setup affects the lift. Recorded native physics; rigid synthetic objects; no physical robot qualification.
Explore the graspability comparison → From reaching a shoe to keeping hold. Compare a fixed grasp sequence with continuous acquisition checks, a measured lift checkpoint and bounded grip-loss recovery. Recorded native MuJoCo attempts preserve the full comparison, failures and independent scoring. Authored policies; synthetic contact; no physical robot qualification.
Inspect the grip comparison and every result → Before a skill can succeed. Explore a hinged shoe box and an out-of-reach shoe. Recorded native MuJoCo attempts retain failed access, withdrawal and tool-assisted reach. Authored policies; synthetic properties; no physical robot qualification.
Inspect both simulations and every result → Native grasp, carry and release.
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.
Hardware profiles and the next experiments. Recorded native contact task 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 ↗ PiPER reliability v0.5 · synthetic Measure placement reliability. Placement completion: v0.4 4/12, v0.5 second lift 6/12, v0.5 full 7/12.
v0.5 second lift versus v0.4: 2 gained, 0 lost, 4 shared successful scenes, 6 neither completed. Mean completion-time change on shared successes: +0.000 simulated seconds (candidate minus reference).
v0.5 full versus v0.4: 3 gained, 0 lost, 4 shared successful scenes, 5 neither completed. Mean completion-time change on shared successes: +0.847 simulated seconds (candidate minus reference).
v0.5 full versus v0.5 second lift: 1 gained, 0 lost, 6 shared successful scenes, 5 neither completed. Mean completion-time change on shared successes: +0.434 simulated seconds (candidate minus reference).
Protocol cohort: confirmation; seed 832847. 84 planned attempts across twelve synthetic scenes. Matched open-jaw and hold controls: 0/48 completed. Every failed or invalid attempt remains in the denominator. Cohorts are reported separately; a cohort label alone does not establish independence. Counts are descriptive synthetic evidence, not physical reliability, detector accuracy, learned competence or real-time sensor performance.
Inspect every stratum, control and pilot gate Pilot gate complete planned denominator: passedThe original height, retention and placement gates remain. Second-lift and alternate-grasp controllers use privileged simulator observations; a recovered grip is not automatically a completed placement. These are recorded receipts, with no PiPER 3D viewer or live browser physics.
PiPER reliability v0.5 · synthetic Measure placement reliability. Placement completion: v0.4 6/12, v0.5 second lift 6/12, v0.5 full 8/12.
v0.5 second lift versus v0.4: 0 gained, 0 lost, 6 shared successful scenes, 6 neither completed. Mean completion-time change on shared successes: +0.000 simulated seconds (candidate minus reference).
v0.5 full versus v0.4: 2 gained, 0 lost, 6 shared successful scenes, 4 neither completed. Mean completion-time change on shared successes: +0.698 simulated seconds (candidate minus reference).
v0.5 full versus v0.5 second lift: 2 gained, 0 lost, 6 shared successful scenes, 4 neither completed. Mean completion-time change on shared successes: +0.698 simulated seconds (candidate minus reference).
Protocol cohort: confirmation; seed 227037. 84 planned attempts across twelve synthetic scenes. Matched open-jaw and hold controls: 0/48 completed. Every failed or invalid attempt remains in the denominator. Cohorts are reported separately; a cohort label alone does not establish independence. Counts are descriptive synthetic evidence, not physical reliability, detector accuracy, learned competence or real-time sensor performance.
Inspect every stratum, control and pilot gate Pilot gate complete planned denominator: passedThe original height, retention and placement gates remain. Second-lift and alternate-grasp controllers use privileged simulator observations; a recovered grip is not automatically a completed placement. These are recorded receipts, with no PiPER 3D viewer or live browser physics.
PiPER reliability v0.5 · synthetic Measure placement reliability. Placement completion: v0.4 5/12, v0.5 second lift 6/12, v0.5 full 8/12.
v0.5 second lift versus v0.4: 1 gained, 0 lost, 5 shared successful scenes, 6 neither completed. Mean completion-time change on shared successes: +0.000 simulated seconds (candidate minus reference).
v0.5 full versus v0.4: 3 gained, 0 lost, 5 shared successful scenes, 4 neither completed. Mean completion-time change on shared successes: +0.800 simulated seconds (candidate minus reference).
v0.5 full versus v0.5 second lift: 2 gained, 0 lost, 6 shared successful scenes, 4 neither completed. Mean completion-time change on shared successes: +0.602 simulated seconds (candidate minus reference).
Protocol cohort: confirmation; seed 862186. 84 planned attempts across twelve synthetic scenes. Matched open-jaw and hold controls: 0/48 completed. Every failed or invalid attempt remains in the denominator. Cohorts are reported separately; a cohort label alone does not establish independence. Counts are descriptive synthetic evidence, not physical reliability, detector accuracy, learned competence or real-time sensor performance.
Inspect every stratum, control and pilot gate Pilot gate complete planned denominator: passedThe original height, retention and placement gates remain. Second-lift and alternate-grasp controllers use privileged simulator observations; a recovered grip is not automatically a completed placement. These are recorded receipts, with no PiPER 3D viewer or live browser physics.
PiPER grip recovery v0.4 · synthetic Measure recovery and grip refusal. Placement completion: v0.3 4/12, v0.4 carry-only 4/12, v0.4 full 4/12.
Retries, resumes and grip refusals do not imply successful placement or recovery. Shared successful scenes: 4; gained placements: 0; lost placements: 0. Mean completion-time change on shared successes: -1.270 simulated seconds (v0.4 minus v0.3). Hold-duration and post-resume motion remain separate in the report.
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.
Inspect all friction, mass and control counts μ=0.80, mass=0.04 kg / history_placement: 2/2 completed, 2 numerically validFrozen v0.3, v0.4 carry-only and v0.4 full use privileged native simulator state. Each placement policy has its own matched open-jaw replay; a shared hold control remains. Failed grip tests, native motion, 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.
PiPER contact memory v0.3 · synthetic Compare probe and contact history. 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.
Inspect all friction, mass and control counts μ=0.80, mass=0.04 kg / baseline_placement: 2/2 completed, 2 numerically validThree 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.
Recorded PiPER contact pilot · synthetic PiPER: authored grasp, carry and release 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.
Frozen confirmation cohort: 9 planned attempts across 3 synthetic scenes. Counts are descriptive, with no population or physical reliability claim.
PiPER guarded grip comparison · synthetic Test the grip. Inspect recovery. 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 valid
μ=0.80 / baseline_open_jaw: 0/2 completed, 2 numerically valid
μ=0.80 / guarded_placement: 2/2 completed, 2 numerically valid
μ=0.80 / guarded_open_jaw: 0/2 completed, 2 numerically valid
μ=0.80 / hold: 0/2 completed, 2 numerically valid
μ=0.55 / baseline_placement: 2/2 completed, 2 numerically valid
μ=0.55 / baseline_open_jaw: 0/2 completed, 2 numerically valid
μ=0.55 / guarded_placement: 0/2 completed, 2 numerically valid
μ=0.55 / guarded_open_jaw: 0/2 completed, 2 numerically valid
μ=0.55 / hold: 0/2 completed, 2 numerically valid
μ=0.15 / baseline_placement: 0/2 completed, 2 numerically valid
μ=0.15 / baseline_open_jaw: 0/2 completed, 2 numerically valid
μ=0.15 / guarded_placement: 0/2 completed, 2 numerically valid
μ=0.15 / guarded_open_jaw: 0/2 completed, 2 numerically valid
μ=0.15 / hold: 0/2 completed, 2 numerically valid
Baseline 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.
Recorded native mechanics · unloaded AgileX PiPER 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 ↗ Recorded native mechanics · unloaded The Robot Studio SO101 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 ↗ Upcoming · no task evidence Two arms and dexterous hands 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.
Run it on your own machine. 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.