unite: cotrain U-Socket + ChainGripper (six-point actions) - #56
Open
aidang3019 wants to merge 4 commits into
Open
unite: cotrain U-Socket + ChainGripper (six-point actions)#56aidang3019 wants to merge 4 commits into
aidang3019 wants to merge 4 commits into
Conversation
ChainGripper joins UNITE as a second embodiment in Elmo's six-point action space [Lx, Ly, Cx, Cy, Rx, Ry] (chain_gripper_control.py at 6bb7b41/bf632fd, dropped from the tree in the 09-04 cleanup). Ported dependency-free as egomimic/rldb/zarr/chain_gripper_points.py (bit-identical FK/IK to the simulator), with the load-time transforms, the points->native decoder and a per-embodiment observation adapter. - model bf/ct_unite_register_separate_nt8_h384_s42: the half-width un-shared recipe with both domains (4-D and 6-D decoders); bf/ch_... chain-only row - data unite_cotrain_usocket_chain_val01_h16 (one batch per embodiment per step) and unite_chain_points_val01_h16_per_emb_proprio - Paper-DP baseline in the same six-point space: bf_planar_v2_dp_paper_points6 + pusht/planar_v2_chain_points_dp_paper - evaluator: semantic_blocks_by_embodiment (energy-score blocks must partition the width, 4 vs 6) - launchers: launch_unite_cotrain.sbatch (per-embodiment dataset roots, manifests, SHAs; FAST keeps throughput settings, not the LR override), launch_planar_chain.sbatch (chain DP; SLURM_RESTART_COUNT=0 preflight fix, quoted ckpt_path), validate_planar_dataset.py --domain - rollout driver: --pusher chain_gripper --chain-control-mode points feeds raw points to the env's point mode; SIM_ROOT selects the simulator tree; per-embodiment decoders from evaluator.native_decoders - tests/test_chain_gripper_points.py Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013EqNns33pX983azhohZXX3
Author
|
Warning This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
This stack of pull requests is managed by Graphite. Learn more about stacking. |
… from the rollout branch Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013EqNns33pX983azhohZXX3
…probe) Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013EqNns33pX983azhohZXX3
…n row Topology B (PerEmbodimentTokenizerUniteGenerativeEncoder): one tokenizer DiT, condition projection, null input, domain embedding and output norm per embodiment; one denoiser DiT shared by both; register latent and per-embodiment decoders unchanged. Built by build_configurable_unite_generative_encoder( per_embodiment_tokenizer=True); model row ct_unite_register_split_tok_nt8_h384_s42 differs from the topology-A row by exactly that flag (asserted in tests). Paper-DP cotrain (planar_v2_cotrain_dp_paper_points6): U-Socket common five zero-padded to six (PadActionWidth / get_planar_paper_padded_transform_list, PaddedPlanarCommon5NativeDecoder) + ChainGripper six points through the existing six-wide Paper-DP head; one batch per embodiment per step, EMA, Hobs 2. launch_unite_cotrain.sbatch now hosts the whole sweep: ICE_FAMILY is derived from the experiment (unite | planar), the Paper-DP BC rows for both embodiments and the DP cotrain row are allowlisted, UNITE-only overrides/asserts are gated on the family. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013EqNns33pX983azhohZXX3
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
ChainGripper joins UNITE as a second embodiment in Elmo's six-point action
space [Lx, Ly, Cx, Cy, Rx, Ry] (chain_gripper_control.py at 6bb7b41/bf632fd,
dropped from the tree in the 09-04 cleanup). Ported dependency-free as
egomimic/rldb/zarr/chain_gripper_points.py (bit-identical FK/IK to the
simulator), with the load-time transforms, the points->native decoder and a
per-embodiment observation adapter.
recipe with both domains (4-D and 6-D decoders); bf/ch_... chain-only row
step) and unite_chain_points_val01_h16_per_emb_proprio
the width, 4 vs 6)
manifests, SHAs; FAST keeps throughput settings, not the LR override),
launch_planar_chain.sbatch (chain DP; SLURM_RESTART_COUNT=0 preflight fix,
quoted ckpt_path), validate_planar_dataset.py --domain
points to the env's point mode; SIM_ROOT selects the simulator tree;
per-embodiment decoders from evaluator.native_decoders
Runs (ICE, 2026-09-08): DP chain points6 300k (job 5739575), UNITE cotrain separate h384 150k (5740150), UNITE chain-only control (5740151). Closed-loop protocol as in #55; the chain gripper is rolled out through the simulator's point mode (
2133a92sim tree). Vault noteUNITE Cotrain Loop 9-8.Stacked on #55.
🤖 Generated with Claude Code
https://claude.ai/code/session_013EqNns33pX983azhohZXX3