λ₀

λ₀ is a humanoid vision-language-action model from Alibaba Token Hub and Amap, Alibaba Group. It learns whole-body loco-manipulation from egocentric human data and predicts motion tokens and hand commands from images, instructions and proprioception. Authors report experiments on Unitree G1 with Revo2 hands.

View full model description

arXiv records v2 on October21;the catalog cites this fixed revision for architecture and evaluation protocols. The revision date does not establish a new model version,a capability breakthrough or a weight/code-license change.

Ordinary history:an officially dated report revision without independent evidence of a material capability or availability change.

Papers & reports (1)

λ₀: Scaling Open-World Egocentric Human Data for Humanoid Loco-Manipulation

Technical report

arXiv: 2610.18950DOI: 10.48550/arxiv.2610.18950
Authors & affiliations (16)

Author order: Team authorship

  1. Lambda-humanoid Team

    Alibaba Token Hub · Amap · Alibaba Group

  2. Chongyang Xucore contributor

    Not provided

  3. Jin Chencore contributor

    Not provided

  4. Jinhui Yecore contributor

    Not provided

  5. Yilun Chencore contributor; project lead

    Not provided

  6. Yiming Jiangcore contributor

    Not provided

  7. Zhao Wucore contributor

    Not provided

  8. Hu Weicontributor

    Not provided

  9. Lin Qucontributor

    Not provided

First published
2026-10-13
Latest paper revision
v2 · 2026-10-21
Venue
Not provided

Covered model versions

  • lambda · λ₀

    λ₀researchmodel andthree-stagepipeline;reportedStageIIIevaluationsareversion-owned,notattributedtothepublicrawStageIImidtraincheckpoint. SIMPLEtask-wiseconfigurationselection andpartial-progressscopesretained.

Paper ↗PDF ↗

How to access and use

Public code resources
Public access
Model files
Public access
Training data
Partly available

Training / fine-tuning code

Access
Public access

SourceVerified on · 2026-11-12

Code license
Apache-2.0

Model files

Model files
λ₀ Midtrain (100k)AccessWeights licenseCommercial use
Allowed under cited termsTerms

SourceVerified on · 2026-11-12

Conditions

Apache-2.0 permits commercial use subject to license/notice obligations for these code and raw Stage-II weights. This is not a task-ready policy or a license for HumanVerse samples,SONIC,other dependencies or a hosted service. Full training data are absent;paperEMA initialization differs from the raw release.

Release information

SourceVerified on · 2026-11-12

Checks before deployment

The available checkpoint contains Stage II 100k raw weights, without EMA or a task-specific post-trained policy. Full training data are unavailable. Stage III and controller integration require further work. Results are author-reported. Available weights or code do not establish plug-and-play robot compatibility. Check the required configuration, data and interfaces in the source documentation.

Deployment and usage resources

Runtime and frameworks
PyTorch
Fine-tuning requirements
Paper stagesI200k/II100k/III50k;publicrawmidtrain differs frompaperEMAinit;completeworkflowsTODO
Deployment requirements
ResearchG1/Revo2+SONIC;Python3.11/3.12 code reference;controllerintegrationincomplete,minimumGPUunknown

Architecture and inputs / outputs

Model type
Action policy
Backbone
Qwen3.5-2B vision-language backbone plus flow-matching action expert
Weight parameters
2.3892 B
Input
Language / RGB images / Robot state
Output
G1:50-step chunks;76D/step=64SONIC latents+two6D hand commands
Released
October 13, 2026

Scope and training

Functional purpose

Robot action control

Training data sources

7ego-hand datasets;HumanVerse500hbeforefiltering;robotposttrainingdata;fulltrainingcorpusnotreleased

Evaluation results

No verified independent benchmark results are recorded for this version or the selected weights.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · Real robot task success and progress ↗

Bottle Disposal · Success: 50 %

arxiv-v2-real-four-tasks

Table1/AppendixD.3:UnitreeG1/Revo2+SONIC;Bottle Disposal,Toy Storage,Box Transport,Chair Placement,ten frozen initial configurations each,40trials/policy. Full autonomous success is separate from100-point milestone progress;humancontact disqualifies autonomous success andterminal condition lasts3seconds. Equal task averages.25/40=62.5% versus80.4%progress. README names a different task set;paper protocol is authoritative. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · Real robot task success and progress ↗

Bottle Disposal · Milestone progress: 79 %

arxiv-v2-real-four-tasks

Table1/AppendixD.3:UnitreeG1/Revo2+SONIC;Bottle Disposal,Toy Storage,Box Transport,Chair Placement,ten frozen initial configurations each,40trials/policy. Full autonomous success is separate from100-point milestone progress;humancontact disqualifies autonomous success andterminal condition lasts3seconds. Equal task averages.25/40=62.5% versus80.4%progress. README names a different task set;paper protocol is authoritative. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · Real robot task success and progress ↗

Toy Storage · Success: 70 %

arxiv-v2-real-four-tasks

Table1/AppendixD.3:UnitreeG1/Revo2+SONIC;Bottle Disposal,Toy Storage,Box Transport,Chair Placement,ten frozen initial configurations each,40trials/policy. Full autonomous success is separate from100-point milestone progress;humancontact disqualifies autonomous success andterminal condition lasts3seconds. Equal task averages.25/40=62.5% versus80.4%progress. README names a different task set;paper protocol is authoritative. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · Real robot task success and progress ↗

Toy Storage · Milestone progress: 81 %

arxiv-v2-real-four-tasks

Table1/AppendixD.3:UnitreeG1/Revo2+SONIC;Bottle Disposal,Toy Storage,Box Transport,Chair Placement,ten frozen initial configurations each,40trials/policy. Full autonomous success is separate from100-point milestone progress;humancontact disqualifies autonomous success andterminal condition lasts3seconds. Equal task averages.25/40=62.5% versus80.4%progress. README names a different task set;paper protocol is authoritative. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · Real robot task success and progress ↗

Box Transport · Success: 70 %

arxiv-v2-real-four-tasks

Table1/AppendixD.3:UnitreeG1/Revo2+SONIC;Bottle Disposal,Toy Storage,Box Transport,Chair Placement,ten frozen initial configurations each,40trials/policy. Full autonomous success is separate from100-point milestone progress;humancontact disqualifies autonomous success andterminal condition lasts3seconds. Equal task averages.25/40=62.5% versus80.4%progress. README names a different task set;paper protocol is authoritative. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · Real robot task success and progress ↗

Box Transport · Milestone progress: 84.5 %

arxiv-v2-real-four-tasks

Table1/AppendixD.3:UnitreeG1/Revo2+SONIC;Bottle Disposal,Toy Storage,Box Transport,Chair Placement,ten frozen initial configurations each,40trials/policy. Full autonomous success is separate from100-point milestone progress;humancontact disqualifies autonomous success andterminal condition lasts3seconds. Equal task averages.25/40=62.5% versus80.4%progress. README names a different task set;paper protocol is authoritative. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · Real robot task success and progress ↗

Chair Placement · Success: 60 %

arxiv-v2-real-four-tasks

Table1/AppendixD.3:UnitreeG1/Revo2+SONIC;Bottle Disposal,Toy Storage,Box Transport,Chair Placement,ten frozen initial configurations each,40trials/policy. Full autonomous success is separate from100-point milestone progress;humancontact disqualifies autonomous success andterminal condition lasts3seconds. Equal task averages.25/40=62.5% versus80.4%progress. README names a different task set;paper protocol is authoritative. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · Real robot task success and progress ↗

Chair Placement · Milestone progress: 77 %

arxiv-v2-real-four-tasks

Table1/AppendixD.3:UnitreeG1/Revo2+SONIC;Bottle Disposal,Toy Storage,Box Transport,Chair Placement,ten frozen initial configurations each,40trials/policy. Full autonomous success is separate from100-point milestone progress;humancontact disqualifies autonomous success andterminal condition lasts3seconds. Equal task averages.25/40=62.5% versus80.4%progress. README names a different task set;paper protocol is authoritative. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · Real robot task success and progress ↗

Mean · Success: 62.5 %

arxiv-v2-real-four-tasks

Table1/AppendixD.3:UnitreeG1/Revo2+SONIC;Bottle Disposal,Toy Storage,Box Transport,Chair Placement,ten frozen initial configurations each,40trials/policy. Full autonomous success is separate from100-point milestone progress;humancontact disqualifies autonomous success andterminal condition lasts3seconds. Equal task averages.25/40=62.5% versus80.4%progress. README names a different task set;paper protocol is authoritative. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · Real robot task success and progress ↗

Mean · Milestone progress: 80.4 %

arxiv-v2-real-four-tasks

Table1/AppendixD.3:UnitreeG1/Revo2+SONIC;Bottle Disposal,Toy Storage,Box Transport,Chair Placement,ten frozen initial configurations each,40trials/policy. Full autonomous success is separate from100-point milestone progress;humancontact disqualifies autonomous success andterminal condition lasts3seconds. Equal task averages.25/40=62.5% versus80.4%progress. README names a different task set;paper protocol is authoritative. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · SIMPLE Level0 ↗

XMovePick L0: 100 %

arxiv-v2-simple-l0

Table2/AppendixD.2:SIMPLE six core tasks×Levels0–2×ten rollouts=180,total162successes. Best recorded configuration per task;no shared checkpoint or inference setting,selection not established as validation-only. Single-task training;not a fixed-configuration comparison. Public baselines retain their own source protocols. Exact evaluated artifacts unreported;not the raw-midtrain download. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · SIMPLE Level0 ↗

BendPick L0: 100 %

arxiv-v2-simple-l0

Table2/AppendixD.2:SIMPLE six core tasks×Levels0–2×ten rollouts=180,total162successes. Best recorded configuration per task;no shared checkpoint or inference setting,selection not established as validation-only. Single-task training;not a fixed-configuration comparison. Public baselines retain their own source protocols. Exact evaluated artifacts unreported;not the raw-midtrain download. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · SIMPLE Level0 ↗

Handover L0: 90 %

arxiv-v2-simple-l0

Table2/AppendixD.2:SIMPLE six core tasks×Levels0–2×ten rollouts=180,total162successes. Best recorded configuration per task;no shared checkpoint or inference setting,selection not established as validation-only. Single-task training;not a fixed-configuration comparison. Public baselines retain their own source protocols. Exact evaluated artifacts unreported;not the raw-midtrain download. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · SIMPLE Level0 ↗

Mobile P&P L0: 90 %

arxiv-v2-simple-l0

Table2/AppendixD.2:SIMPLE six core tasks×Levels0–2×ten rollouts=180,total162successes. Best recorded configuration per task;no shared checkpoint or inference setting,selection not established as validation-only. Single-task training;not a fixed-configuration comparison. Public baselines retain their own source protocols. Exact evaluated artifacts unreported;not the raw-midtrain download. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · SIMPLE Level0 ↗

Grasp L0: 100 %

arxiv-v2-simple-l0

Table2/AppendixD.2:SIMPLE six core tasks×Levels0–2×ten rollouts=180,total162successes. Best recorded configuration per task;no shared checkpoint or inference setting,selection not established as validation-only. Single-task training;not a fixed-configuration comparison. Public baselines retain their own source protocols. Exact evaluated artifacts unreported;not the raw-midtrain download. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · SIMPLE Level0 ↗

XMoveBendPick L0: 60 %

arxiv-v2-simple-l0

Table2/AppendixD.2:SIMPLE six core tasks×Levels0–2×ten rollouts=180,total162successes. Best recorded configuration per task;no shared checkpoint or inference setting,selection not established as validation-only. Single-task training;not a fixed-configuration comparison. Public baselines retain their own source protocols. Exact evaluated artifacts unreported;not the raw-midtrain download. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · SIMPLE Level0 ↗

Mean: 90 %

arxiv-v2-simple-l0

Table2/AppendixD.2:SIMPLE six core tasks×Levels0–2×ten rollouts=180,total162successes. Best recorded configuration per task;no shared checkpoint or inference setting,selection not established as validation-only. Single-task training;not a fixed-configuration comparison. Public baselines retain their own source protocols. Exact evaluated artifacts unreported;not the raw-midtrain download. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · SIMPLE Level1 ↗

XMovePick L1: 100 %

arxiv-v2-simple-l1

Table2/AppendixD.2:SIMPLE six core tasks×Levels0–2×ten rollouts=180,total162successes. Best recorded configuration per task;no shared checkpoint or inference setting,selection not established as validation-only. Single-task training;not a fixed-configuration comparison. Public baselines retain their own source protocols. Exact evaluated artifacts unreported;not the raw-midtrain download. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · SIMPLE Level1 ↗

BendPick L1: 100 %

arxiv-v2-simple-l1

Table2/AppendixD.2:SIMPLE six core tasks×Levels0–2×ten rollouts=180,total162successes. Best recorded configuration per task;no shared checkpoint or inference setting,selection not established as validation-only. Single-task training;not a fixed-configuration comparison. Public baselines retain their own source protocols. Exact evaluated artifacts unreported;not the raw-midtrain download. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · SIMPLE Level1 ↗

Handover L1: 100 %

arxiv-v2-simple-l1

Table2/AppendixD.2:SIMPLE six core tasks×Levels0–2×ten rollouts=180,total162successes. Best recorded configuration per task;no shared checkpoint or inference setting,selection not established as validation-only. Single-task training;not a fixed-configuration comparison. Public baselines retain their own source protocols. Exact evaluated artifacts unreported;not the raw-midtrain download. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · SIMPLE Level1 ↗

Mobile P&P L1: 70 %

arxiv-v2-simple-l1

Table2/AppendixD.2:SIMPLE six core tasks×Levels0–2×ten rollouts=180,total162successes. Best recorded configuration per task;no shared checkpoint or inference setting,selection not established as validation-only. Single-task training;not a fixed-configuration comparison. Public baselines retain their own source protocols. Exact evaluated artifacts unreported;not the raw-midtrain download. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · SIMPLE Level1 ↗

Grasp L1: 90 %

arxiv-v2-simple-l1

Table2/AppendixD.2:SIMPLE six core tasks×Levels0–2×ten rollouts=180,total162successes. Best recorded configuration per task;no shared checkpoint or inference setting,selection not established as validation-only. Single-task training;not a fixed-configuration comparison. Public baselines retain their own source protocols. Exact evaluated artifacts unreported;not the raw-midtrain download. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · SIMPLE Level1 ↗

XMoveBendPick L1: 100 %

arxiv-v2-simple-l1

Table2/AppendixD.2:SIMPLE six core tasks×Levels0–2×ten rollouts=180,total162successes. Best recorded configuration per task;no shared checkpoint or inference setting,selection not established as validation-only. Single-task training;not a fixed-configuration comparison. Public baselines retain their own source protocols. Exact evaluated artifacts unreported;not the raw-midtrain download. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · SIMPLE Level1 ↗

Mean: 93.3 %

arxiv-v2-simple-l1

Table2/AppendixD.2:SIMPLE six core tasks×Levels0–2×ten rollouts=180,total162successes. Best recorded configuration per task;no shared checkpoint or inference setting,selection not established as validation-only. Single-task training;not a fixed-configuration comparison. Public baselines retain their own source protocols. Exact evaluated artifacts unreported;not the raw-midtrain download. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · SIMPLE Level2 ↗

XMovePick L2: 80 %

arxiv-v2-simple-l2

Table2/AppendixD.2:SIMPLE six core tasks×Levels0–2×ten rollouts=180,total162successes. Best recorded configuration per task;no shared checkpoint or inference setting,selection not established as validation-only. Single-task training;not a fixed-configuration comparison. Public baselines retain their own source protocols. Exact evaluated artifacts unreported;not the raw-midtrain download. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · SIMPLE Level2 ↗

BendPick L2: 100 %

arxiv-v2-simple-l2

Table2/AppendixD.2:SIMPLE six core tasks×Levels0–2×ten rollouts=180,total162successes. Best recorded configuration per task;no shared checkpoint or inference setting,selection not established as validation-only. Single-task training;not a fixed-configuration comparison. Public baselines retain their own source protocols. Exact evaluated artifacts unreported;not the raw-midtrain download. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · SIMPLE Level2 ↗

Handover L2: 100 %

arxiv-v2-simple-l2

Table2/AppendixD.2:SIMPLE six core tasks×Levels0–2×ten rollouts=180,total162successes. Best recorded configuration per task;no shared checkpoint or inference setting,selection not established as validation-only. Single-task training;not a fixed-configuration comparison. Public baselines retain their own source protocols. Exact evaluated artifacts unreported;not the raw-midtrain download. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · SIMPLE Level2 ↗

Mobile P&P L2: 70 %

arxiv-v2-simple-l2

Table2/AppendixD.2:SIMPLE six core tasks×Levels0–2×ten rollouts=180,total162successes. Best recorded configuration per task;no shared checkpoint or inference setting,selection not established as validation-only. Single-task training;not a fixed-configuration comparison. Public baselines retain their own source protocols. Exact evaluated artifacts unreported;not the raw-midtrain download. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · SIMPLE Level2 ↗

Grasp L2: 80 %

arxiv-v2-simple-l2

Table2/AppendixD.2:SIMPLE six core tasks×Levels0–2×ten rollouts=180,total162successes. Best recorded configuration per task;no shared checkpoint or inference setting,selection not established as validation-only. Single-task training;not a fixed-configuration comparison. Public baselines retain their own source protocols. Exact evaluated artifacts unreported;not the raw-midtrain download. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · SIMPLE Level2 ↗

XMoveBendPick L2: 90 %

arxiv-v2-simple-l2

Table2/AppendixD.2:SIMPLE six core tasks×Levels0–2×ten rollouts=180,total162successes. Best recorded configuration per task;no shared checkpoint or inference setting,selection not established as validation-only. Single-task training;not a fixed-configuration comparison. Public baselines retain their own source protocols. Exact evaluated artifacts unreported;not the raw-midtrain download. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · SIMPLE Level2 ↗

Mean: 86.7 %

arxiv-v2-simple-l2

Table2/AppendixD.2:SIMPLE six core tasks×Levels0–2×ten rollouts=180,total162successes. Best recorded configuration per task;no shared checkpoint or inference setting,selection not established as validation-only. Single-task training;not a fixed-configuration comparison. Public baselines retain their own source protocols. Exact evaluated artifacts unreported;not the raw-midtrain download. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · SIMPLE task-wise selected overall ↗

Success: 90 %

arxiv-v2-simple-overall

Table2/AppendixD.2:SIMPLE six core tasks×Levels0–2×ten rollouts=180,total162successes. Best recorded configuration per task;no shared checkpoint or inference setting,selection not established as validation-only. Single-task training;not a fixed-configuration comparison. Public baselines retain their own source protocols. Exact evaluated artifacts unreported;not the raw-midtrain download. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · HumanVerse fraction versus real progress ↗

0% HumanVerse: 59.9 %

arxiv-v2-human-data-scaling

Figure12/AppendixC:Stage-II HumanVerse training fractions0/5/10/25/50/100%;session-disjoint nested subsets with fixedStageI/modelcapacity/robotdata/StageIIIbudget/evaluationmanifest. Nonzero fractions shareStageIIupdatebudget;0omitsStageII. Weighted real-task progress,not binarysuccess;80.4versus59.9is20.5percentagepoints. One-runoffline lossfits do not establish seedreproducibility or extrapolation. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · HumanVerse fraction versus real progress ↗

5% HumanVerse: 64.2 %

arxiv-v2-human-data-scaling

Figure12/AppendixC:Stage-II HumanVerse training fractions0/5/10/25/50/100%;session-disjoint nested subsets with fixedStageI/modelcapacity/robotdata/StageIIIbudget/evaluationmanifest. Nonzero fractions shareStageIIupdatebudget;0omitsStageII. Weighted real-task progress,not binarysuccess;80.4versus59.9is20.5percentagepoints. One-runoffline lossfits do not establish seedreproducibility or extrapolation. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · HumanVerse fraction versus real progress ↗

10% HumanVerse: 67.3 %

arxiv-v2-human-data-scaling

Figure12/AppendixC:Stage-II HumanVerse training fractions0/5/10/25/50/100%;session-disjoint nested subsets with fixedStageI/modelcapacity/robotdata/StageIIIbudget/evaluationmanifest. Nonzero fractions shareStageIIupdatebudget;0omitsStageII. Weighted real-task progress,not binarysuccess;80.4versus59.9is20.5percentagepoints. One-runoffline lossfits do not establish seedreproducibility or extrapolation. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · HumanVerse fraction versus real progress ↗

25% HumanVerse: 72.1 %

arxiv-v2-human-data-scaling

Figure12/AppendixC:Stage-II HumanVerse training fractions0/5/10/25/50/100%;session-disjoint nested subsets with fixedStageI/modelcapacity/robotdata/StageIIIbudget/evaluationmanifest. Nonzero fractions shareStageIIupdatebudget;0omitsStageII. Weighted real-task progress,not binarysuccess;80.4versus59.9is20.5percentagepoints. One-runoffline lossfits do not establish seedreproducibility or extrapolation. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · HumanVerse fraction versus real progress ↗

50% HumanVerse: 76.4 %

arxiv-v2-human-data-scaling

Figure12/AppendixC:Stage-II HumanVerse training fractions0/5/10/25/50/100%;session-disjoint nested subsets with fixedStageI/modelcapacity/robotdata/StageIIIbudget/evaluationmanifest. Nonzero fractions shareStageIIupdatebudget;0omitsStageII. Weighted real-task progress,not binarysuccess;80.4versus59.9is20.5percentagepoints. One-runoffline lossfits do not establish seedreproducibility or extrapolation. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · HumanVerse fraction versus real progress ↗

100% HumanVerse: 80.4 %

arxiv-v2-human-data-scaling

Figure12/AppendixC:Stage-II HumanVerse training fractions0/5/10/25/50/100%;session-disjoint nested subsets with fixedStageI/modelcapacity/robotdata/StageIIIbudget/evaluationmanifest. Nonzero fractions shareStageIIupdatebudget;0omitsStageII. Weighted real-task progress,not binarysuccess;80.4versus59.9is20.5percentagepoints. One-runoffline lossfits do not establish seedreproducibility or extrapolation. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · No human-data stages ↗

SIMPLE overall: 79.4 %

arxiv-v2-no-human-stages

Table3 ablation No human-data stages:all variants retain architecture/action interfaces/VLM initialization andsameStageIIIrobotgrounding. Real metrics use fourtasks×ten trials,success versusmilestoneprogress. SIMPLEisreported task-wiseconfigurationselection;matchedcheckpoint/inferencesettingsacrossablationsareunverified. Do notdescribe these scores as an isolated current-download checkpoint test. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · No human-data stages ↗

Real success: 15 %

arxiv-v2-no-human-stages

Table3 ablation No human-data stages:all variants retain architecture/action interfaces/VLM initialization andsameStageIIIrobotgrounding. Real metrics use fourtasks×ten trials,success versusmilestoneprogress. SIMPLEisreported task-wiseconfigurationselection;matchedcheckpoint/inferencesettingsacrossablationsareunverified. Do notdescribe these scores as an isolated current-download checkpoint test. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · No human-data stages ↗

Milestone progress: 54 %

arxiv-v2-no-human-stages

Table3 ablation No human-data stages:all variants retain architecture/action interfaces/VLM initialization andsameStageIIIrobotgrounding. Real metrics use fourtasks×ten trials,success versusmilestoneprogress. SIMPLEisreported task-wiseconfigurationselection;matchedcheckpoint/inferencesettingsacrossablationsareunverified. Do notdescribe these scores as an isolated current-download checkpoint test. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · Without Stage II ↗

SIMPLE overall: 82.3 %

arxiv-v2-no-stage-ii

Table3 ablation Without Stage II:all variants retain architecture/action interfaces/VLM initialization andsameStageIIIrobotgrounding. Real metrics use fourtasks×ten trials,success versusmilestoneprogress. SIMPLEisreported task-wiseconfigurationselection;matchedcheckpoint/inferencesettingsacrossablationsareunverified. Do notdescribe these scores as an isolated current-download checkpoint test. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · Without Stage II ↗

Real success: 22.5 %

arxiv-v2-no-stage-ii

Table3 ablation Without Stage II:all variants retain architecture/action interfaces/VLM initialization andsameStageIIIrobotgrounding. Real metrics use fourtasks×ten trials,success versusmilestoneprogress. SIMPLEisreported task-wiseconfigurationselection;matchedcheckpoint/inferencesettingsacrossablationsareunverified. Do notdescribe these scores as an isolated current-download checkpoint test. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · Without Stage II ↗

Milestone progress: 59.9 %

arxiv-v2-no-stage-ii

Table3 ablation Without Stage II:all variants retain architecture/action interfaces/VLM initialization andsameStageIIIrobotgrounding. Real metrics use fourtasks×ten trials,success versusmilestoneprogress. SIMPLEisreported task-wiseconfigurationselection;matchedcheckpoint/inferencesettingsacrossablationsareunverified. Do notdescribe these scores as an isolated current-download checkpoint test. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · Without Stage I ↗

SIMPLE overall: 87.8 %

arxiv-v2-no-stage-i

Table3 ablation Without Stage I:all variants retain architecture/action interfaces/VLM initialization andsameStageIIIrobotgrounding. Real metrics use fourtasks×ten trials,success versusmilestoneprogress. SIMPLEisreported task-wiseconfigurationselection;matchedcheckpoint/inferencesettingsacrossablationsareunverified. Do notdescribe these scores as an isolated current-download checkpoint test. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · Without Stage I ↗

Real success: 50 %

arxiv-v2-no-stage-i

Table3 ablation Without Stage I:all variants retain architecture/action interfaces/VLM initialization andsameStageIIIrobotgrounding. Real metrics use fourtasks×ten trials,success versusmilestoneprogress. SIMPLEisreported task-wiseconfigurationselection;matchedcheckpoint/inferencesettingsacrossablationsareunverified. Do notdescribe these scores as an isolated current-download checkpoint test. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · Without Stage I ↗

Milestone progress: 75.9 %

arxiv-v2-no-stage-i

Table3 ablation Without Stage I:all variants retain architecture/action interfaces/VLM initialization andsameStageIIIrobotgrounding. Real metrics use fourtasks×ten trials,success versusmilestoneprogress. SIMPLEisreported task-wiseconfigurationselection;matchedcheckpoint/inferencesettingsacrossablationsareunverified. Do notdescribe these scores as an isolated current-download checkpoint test. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · Robot-unseen,human-observed objects ↗

Human-guided progress: 55.8 %

arxiv-v2-robot-unseen-human-observed

Section5.4/Figure13/AppendixD.3:robot demonstrations containobjectsA;human-guided recipecontains humanA+B. SamepolicytestedonA/B;robot-onlyvarianttestedonB. Ten trials per task/method/condition overfourtasks. B is robot-unseen buthuman-observed,notfullyunseen-object zero-shot. Human-guidedmean55.8 versusrobot-only48.4progress;combinedhumantrainingbenefit doesnotisolatehumanBversushumanA/earlierpretraining. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Publisher evaluation · Reported; not independently reproduced

λ₀ official reported results · Robot-unseen,human-observed objects ↗

Robot-only progress: 48.4 %

arxiv-v2-robot-unseen-human-observed

Section5.4/Figure13/AppendixD.3:robot demonstrations containobjectsA;human-guided recipecontains humanA+B. SamepolicytestedonA/B;robot-onlyvarianttestedonB. Ten trials per task/method/condition overfourtasks. B is robot-unseen buthuman-observed,notfullyunseen-object zero-shot. Human-guidedmean55.8 versusrobot-only48.4progress;combinedhumantrainingbenefit doesnotisolatehumanBversushumanA/earlierpretraining. Developer-reported v2 results,not independent replication. Evaluated immutable weights and run dates are unreported. Version-owned;the public raw Stage-II artifact is not the evaluated Stage-III policy.

Full parameters and verification notes (model card)

https://arxiv.org/abs/2610.18950 https://lambdahumanoid.github.io/ https://github.com/LambdaHumanoid/openlambda https://huggingface.co/lambda-humanoid/lambda0-midtrain-100k

The public checkpoint is Stage II after100,000 updates:raw weights without EMA or a task-specific Stage-III policy. The paper initializes Stage III from Stage-II EMA;the current example uses raw weights,so it is not an exact released reproduction. Current code includes example post-training configs,but complete Stage-III evaluation and robot-controller deployment remain TODO. The older model card says that recipe is being prepared. The stored2.3892B count is rounded from2,389,229,762 tensor elements,independently counted from the safetensors header:2,213,241,664BF16 and175,988,098FP32;the2B backbone label is not the full count. No tensor payload was executed. G1 uses44D state and76D action per step:64SONIC latents plus12hand commands,with50-step chunks and10flow steps. This does not establish50Hz inference. Stage-I138D hand geometry,other human interfaces and36D SIMPLE actions are different contracts. HumanVerse500h/32,993episodes/87million pose frames precede conversion and filtering;the full corpus is not public. The viewer only exposes20eight-second clips with unspecified dataset license. Apache2applies to the reviewed code/weights,not all data,SONIC or other dependencies. No provider-hosted API terms are verified. Paper-v2 real tasks are Bottle Disposal,Toy Storage,Box Transport and Chair Placement,ten trials each:25/40complete successes,62.5%,versus80.4%milestone progress. The repository names other tasks in that summary;use the paper protocol. SIMPLE90%=162/180selects the best configuration per task,without one shared checkpoint/inference setting or proven validation-only selection;matched ablation selection is unverified. Alignment-pipeline51.7/76.2%are partial progress without target-task robot data,not λ-policy binary success. Robot-unseen B objects are human-observed;55.8%progress is not fully unseen-object zero-shot. All evaluations are developer reports;immutable evaluated weights/run dates are unknown. The mentioned SIMPLE checkpoint requires original assets and is not publicly accessible in the current registry;no second checkpoint is invented.

Submit a source or correction