No verified independent benchmark results are recorded for this version or the selected weights.
Publisher evaluation · Reported; not independently reproduced
LaWAM original paper v2 · LIBERO Long ↗Success rate: 97 %
Original LaWAM paper v2 developer-reported research results, version-scoped rather than assigned to a downloaded artifact or LeRobot conversion. Protocols and latency overhead exclusions remain explicit. Table1/AppendixF.2:40LIBERO tasks,50trials/task,2000trials total after benchmark-specific25k-step post-training. Mean of four equally sized10-task suites. Distinct from LeRobot native-format1970/2000=98.5% and MuJoCo3.3.2 evaluation.
Publisher evaluation · Reported; not independently reproduced
LaWAM original paper v2 · LIBERO Goal ↗Success rate: 98.4 %
Original LaWAM paper v2 developer-reported research results, version-scoped rather than assigned to a downloaded artifact or LeRobot conversion. Protocols and latency overhead exclusions remain explicit. Table1/AppendixF.2:40LIBERO tasks,50trials/task,2000trials total after benchmark-specific25k-step post-training. Mean of four equally sized10-task suites. Distinct from LeRobot native-format1970/2000=98.5% and MuJoCo3.3.2 evaluation.
Publisher evaluation · Reported; not independently reproduced
LaWAM original paper v2 · LIBERO Object ↗Success rate: 99.6 %
Original LaWAM paper v2 developer-reported research results, version-scoped rather than assigned to a downloaded artifact or LeRobot conversion. Protocols and latency overhead exclusions remain explicit. Table1/AppendixF.2:40LIBERO tasks,50trials/task,2000trials total after benchmark-specific25k-step post-training. Mean of four equally sized10-task suites. Distinct from LeRobot native-format1970/2000=98.5% and MuJoCo3.3.2 evaluation.
Publisher evaluation · Reported; not independently reproduced
LaWAM original paper v2 · LIBERO Spatial ↗Success rate: 99.4 %
Original LaWAM paper v2 developer-reported research results, version-scoped rather than assigned to a downloaded artifact or LeRobot conversion. Protocols and latency overhead exclusions remain explicit. Table1/AppendixF.2:40LIBERO tasks,50trials/task,2000trials total after benchmark-specific25k-step post-training. Mean of four equally sized10-task suites. Distinct from LeRobot native-format1970/2000=98.5% and MuJoCo3.3.2 evaluation.
Publisher evaluation · Reported; not independently reproduced
LaWAM original paper v2 · LIBERO Average ↗Success rate: 98.6 %
Original LaWAM paper v2 developer-reported research results, version-scoped rather than assigned to a downloaded artifact or LeRobot conversion. Protocols and latency overhead exclusions remain explicit. Table1/AppendixF.2:40LIBERO tasks,50trials/task,2000trials total after benchmark-specific25k-step post-training. Mean of four equally sized10-task suites. Distinct from LeRobot native-format1970/2000=98.5% and MuJoCo3.3.2 evaluation.
Publisher evaluation · Reported; not independently reproduced
LaWAM original paper v2 · RoboTwin Clean ↗Success rate: 92.64 %
Original LaWAM paper v2 developer-reported research results, version-scoped rather than assigned to a downloaded artifact or LeRobot conversion. Protocols and latency overhead exclusions remain explicit. Table2/AppendixF.3:50RoboTwin2.0 tasks,100trials/task per clean/randomized condition,5000trials/condition. Post-trained100ksteps on2500clean+25000randomized demonstrations. Overall91.22 is the equally weighted mean of clean92.64 and randomized89.80.
Publisher evaluation · Reported; not independently reproduced
LaWAM original paper v2 · RoboTwin Randomized ↗Success rate: 89.8 %
Original LaWAM paper v2 developer-reported research results, version-scoped rather than assigned to a downloaded artifact or LeRobot conversion. Protocols and latency overhead exclusions remain explicit. Table2/AppendixF.3:50RoboTwin2.0 tasks,100trials/task per clean/randomized condition,5000trials/condition. Post-trained100ksteps on2500clean+25000randomized demonstrations. Overall91.22 is the equally weighted mean of clean92.64 and randomized89.80.
Publisher evaluation · Reported; not independently reproduced
LaWAM original paper v2 · RoboTwin Overall ↗Success rate: 91.22 %
Original LaWAM paper v2 developer-reported research results, version-scoped rather than assigned to a downloaded artifact or LeRobot conversion. Protocols and latency overhead exclusions remain explicit. Table2/AppendixF.3:50RoboTwin2.0 tasks,100trials/task per clean/randomized condition,5000trials/condition. Post-trained100ksteps on2500clean+25000randomized demonstrations. Overall91.22 is the equally weighted mean of clean92.64 and randomized89.80.
Publisher evaluation · Reported; not independently reproduced
LaWAM original paper v2 · A100 model-only chunk latency ↗Model-only latency: 187 ms
Original LaWAM paper v2 developer-reported research results, version-scoped rather than assigned to a downloaded artifact or LeRobot conversion. Protocols and latency overhead exclusions remain explicit. AppendixF.1:mean1000CUDA-synchronized model-only action-chunk predictions onA100,BF16,10action-expert denoising steps. Excludes simulator/robot execution; each baseline retains native resolution/horizon/workload.187ms is not robot control frequency or universal end-to-end latency;24x refers to compared pixel-space WAMs.
