Donghyun Shin

Complete 5,000-update checkpoint, including Orbax params, training state and normalization assets under 4999/. Use openpi training config pi05liberoidmlorareadoutviewadapt with checkpoint directory 4999/. Source jointly trains IL and IDM with global batch64 for each loss. The shared LoRA factors learn from both losses; the identity plus rank16 policy readout learns from IL. Adaptation starts independently from the same source checkpoint, trains shared LoRA with IDM only, and freezes the policy readout and base weights. No KD is used. The published policy uses the readout during inference. Full implementation and experiment provenance are recorded in provenance.json. Pin the verified Hugging…

Matched KD ablation: H10 cumulative-average shared-decoder IDM, IL/KD weights 0, 10,000 updates, global IDM batch 32, LR 1e-5, seed 42, 4-GPU FSDP. View adapts vision/base LLM with a frozen action expert; yaw adapts the full policy. All conditions start independently from the 2026-09-26 source step 9999. MuJoCo 3.2.3. Policy config: pi05liberodensecontact20261002yawidmonly. Data commits and exact commands are in trainingconfiguration.json. Orbax parameters and normalization assets are included; optimizer states remain local.

Matched KD ablation: H10 cumulative-average shared-decoder IDM, IL/KD weights 0, 10,000 updates, global IDM batch 32, LR 1e-5, seed 42, 4-GPU FSDP. View adapts vision/base LLM with a frozen action expert; yaw adapts the full policy. All conditions start independently from the 2026-09-26 source step 9999. MuJoCo 3.2.3. Policy config: pi05liberodensecontact20261002yawidmonly. Data commits and exact commands are in trainingconfiguration.json. Orbax parameters and normalization assets are included; optimizer states remain local.

Matched KD ablation: H10 cumulative-average shared-decoder IDM, IL/KD weights 0, 10,000 updates, global IDM batch 32, LR 1e-5, seed 42, 4-GPU FSDP. View adapts vision/base LLM with a frozen action expert; yaw adapts the full policy. All conditions start independently from the 2026-09-26 source step 9999. MuJoCo 3.2.3. Policy config: pi05liberodensecontact20261002yawidmonly. Data commits and exact commands are in trainingconfiguration.json. Orbax parameters and normalization assets are included; optimizer states remain local.

Raw step999 after 1000 target-only IDM updates, batch32. This pilot trains vision and the base LLM (9 leaves); the graph is gemma300m with action LoRA. Action updates are multiplied by 0.1 after native AdamW. The original source wrist is restored for deployment. The same recipe is used across view and action-frame shifts. No source observations, paired teacher cache, successful task rollout, or task-success feedback is used. Newly unfrozen source parameters are FP32 during training; they were frozen BF16 in the earlier LoRA-only control. Thus this ablation changes trainability and training precision. Base LLM9 initialization is bound to independently restored canonical FP32 source hashes…

Raw step999 after 1000 target-only IDM updates, batch32. This pilot trains vision and the full action expert (11 leaves); the graph is original gemma300m with no LoRA. Action updates are multiplied by 0.1 after native AdamW. The original source wrist is restored for deployment. The same recipe is used across view and action-frame shifts. No source observations, paired teacher cache, successful task rollout, or task-success feedback is used. Newly unfrozen source parameters are FP32 during training; they were frozen BF16 in the earlier LoRA-only control. Thus this ablation changes trainability and training precision. Full-expert initialization verifies loaded-to-initialized bytes and BF16…

Raw step999 after 1000 target-only IDM updates, batch32. This pilot trains vision and the full action expert (11 leaves); the graph is original gemma300m with no LoRA. Action updates are multiplied by 0.1 after native AdamW. The original source wrist is restored for deployment. The same recipe is used across view and action-frame shifts. No source observations, paired teacher cache, successful task rollout, or task-success feedback is used. Newly unfrozen source parameters are FP32 during training; they were frozen BF16 in the earlier LoRA-only control. Thus this ablation changes trainability and training precision. Full-expert initialization verifies loaded-to-initialized bytes and BF16…

Completed source-fresh1000-update target-only IDM pilot, B32, seed42/draw seed170013. Raw999/non-EMA: native51 leaves,43 original-source FP32 trainable leaves and8 frozen projections. Original source-wrist23 leaves are appended for inference; deployment74 restores as BF16. Weights, source normalization, exact controller/source-bank/plan/journal and strict export proof are included. Native full Adam/EMA states remain preserved at the training server. All10 logical collection lineages per suite (40 total) are eligible; eight windows per suite are drawn without replacement each update. All1000 draws and initialization match the full-probe fixed1:1:1 control; the first25 native metrics match…

Completed source-fresh1000-update target-only IDM pilot, B32, seed42/draw seed170013. Raw999/non-EMA: native51 leaves,43 original-source FP32 trainable leaves and8 frozen projections. Original source-wrist23 leaves are appended for inference; deployment74 restores as BF16. Weights, source normalization, exact controller/source-bank/plan/journal and strict export proof are included. Native full Adam/EMA states remain preserved at the training server. All10 logical collection lineages per suite (40 total) are eligible; eight windows per suite are drawn without replacement each update. All1000 draws and initialization match the full-probe fixed1:1:1 control; the first25 native metrics match…

Completed source-fresh1000-update target-only IDM pilot, B32, seed42/draw seed170013. Raw999/non-EMA: native51 leaves,43 original-source FP32 trainable leaves and8 frozen projections. Original source-wrist23 leaves are appended for inference; deployment74 restores as BF16. Weights, source normalization, exact controller/source-bank/plan/journal and strict export proof are included. Native full Adam/EMA states remain preserved at the training server. All10 logical collection lineages per suite (40 total) are eligible; eight windows per suite are drawn without replacement each update. All1000 draws and initialization match the full-probe fixed1:1:1 control; the first25 native metrics match…

viewsmall; 5000 completed updates/raw step 4999, from one continuous run of 5000 updates. LLM frozen; post-Adam update multipliers: vision 1, expert LoRA multiplier 0.1. Existing dense-contact data and original pre-adapt, B32, training seed42, FSDP2, warmup500, peak/end1e-5, IDM-only, raw non-EMA and original source wrist. Native training revision: 622c8d28a737f2e18f910a8e8774499abe11786a. Overall 90.05% (1801/2000); runtime errors0. Official indices10–49: 89.750% (1436/1600). 40 tasks x50 official states0–49, envseed7, A1000–3, eight clients/GPU, MuJoCo3.2.3, native inference revision 0f6471d99dfc81fabe7c3ab65c141d7640b9ff08. All failures retained; one training seed, shared policy RNG…

Fresh matched source Donghyun1228/pi05-libero-support-probe-source-n300000-10000-20261015 @ 2e35994a62b96d29398b7ef137f915afe3d4a40a. Dataset Donghyun1228/libero-support-random-waypoint-300k-20261015 @ 32ed215baf01af7e82f927a62b108873aa70bd19. Only the nine base-LLM leaves are frozen. Vision23, full expert11 and action/time projections8 train with the same native AdamW update multiplier1. B32, seed42, LR1e-5, warmup500, 5000 total updates. Gripper coordinate6 is excluded from target IDM loss (mean of remaining31 padded coordinates); seven-dimensional inputs/output and gripper control remain. Original normalization/source wrist preserved. Raw non-EMA step999; BF16 inference. Evaluation…

Fresh matched source Donghyun1228/pi05-libero-support-probe-source-n300000-10000-20261015 @ 2e35994a62b96d29398b7ef137f915afe3d4a40a. Dataset Donghyun1228/libero-support-random-waypoint-300k-20261015 @ 32ed215baf01af7e82f927a62b108873aa70bd19. Only the nine base-LLM leaves are frozen. Vision23, full expert11 and action/time projections8 train with the same native AdamW update multiplier1. B32, seed42, LR1e-5, warmup500, 5000 total updates. Gripper coordinate6 is excluded from target IDM loss (mean of remaining31 padded coordinates); seven-dimensional inputs/output and gripper control remain. Original normalization/source wrist preserved. Raw non-EMA step2999; BF16 inference. Evaluation…

50k support-near random-probe frames,10000 updates,batch32,seed42,LR1e-5,EMA off. LLM frozen. Source IL trains block32 proper rotations and full non-LLM native weights; IDM bypasses rotations and trains full non-LLM native weights. Targets restore this new source,freeze rotations/anchors and exclude gripper6 from IDM loss. Source warmup100,target warmup500. Replay at0/every1000 uses fixed expert labels only for diagnostics. Top-level checkpoint retains raw weights,anchors,rotations and full optimizer for resume; policy/ contains materialized inference weights and normalization. Target policy wrist is the new materialized source wrist. Source has no simulator success claim. Dataset…

The rollouts of Donghyun1228/libero-policy-rollouts-success-20261002 (successful episodes) and Donghyun1228/libero-policy-rollouts-pooled-20261002 (all episodes) re-packed as one LeRobot v2.1 dataset per camera view, so a single repoid can feed the policy (imitation) loss. Rows are unchanged; the four suites are concatenated with episode, frame and task indices renumbered. meta/episodes.jsonl keeps source: {suite, episodeindex} back to the per-suite datasets. The IDM windows stay in the source repos' /lerobot/alignedidmh10.json manifests and refer to the same episodes. the policy saw only the wrist camera (external camera masked at inference); stored images are always the shifted external…

Independent yaw 10/20/30-degree conditions, original agentview, robot-base yaw compensation and action rotation enabled. MuJoCo 3.2.3. Each buffer has 400 source-policy rollouts across 40 tasks and initial states 0-9, retaining both outcomes. Prompt-only adaptation uses 1,000 steps, batch 128, AdamW LR 1e-5. Source SG-LP training used 20,000 steps at LR 1e-5. Dataset NPZs contain normalized observation/state frames and success metadata, with no expert or predicted-action imitation targets. Normalization is the source policy's fixed 8-D state / 7-D action quantile statistics. Models include full source/adapted weights, source normalization, prompt resume states, protocol/configuration, and…

50,000 frames across40 tasks at20Hz; each task uses the first1,250 frames of the pinned source. 49,600 H10 windows, with no hidden exclusions. MuJoCo3.2.3. Source actions are physically replayed at yaw0, then the simulator state is transferred to compensated yaw20. The external camera is agentviewmedium; wrist images and7D actions retain their normal schema. Saved translation/rotation commands are rotated by -20degrees; gripper is unchanged. This dataset follows the same state-rerender protocol as the existing yaw datasets. Source: Donghyun1228/libero-support-random-waypoint-300k-20261015 at 32ed215baf01af7e82f927a62b108873aa70bd19. Four LeRobot parts are under /lerobot/medium. Use an…