rrp checkpoints: structured latent packets for cross-body robot control
0
4 commits
1 linked in READMEs
updated Oct 4, 2026
Research checkpoints from rrp. Code, decisions and evidence live at
JacobFV/structured-psi0-latent-diffusion-dynamics
(local name relational-robot-policy). Architecture: system i is a flow model that generates a structured latent packet
z[knots × assemblies × 64], supervised by semantic probes. System 0 (our realizer) turns z plus morphology and
proprioception into joint commands; legged bodies also use a learned gait tracker. "System 0" is unrelated to the upstream Ψ₀
model (see docs/related_repos.md in the code repo).
Everything here was trained in MuJoCo simulation by this project. These are research artifacts, not deployable
controllers: no real robot has run them. Sources are labelled exactly as in the decisions log: learned:<ckpt> for our
models, bc:<ckpt> for behaviour-cloning baselines, learned_tracker:<body>:<ver> for gait trackers. Scripted teachers are not
included (they are code).
artifacts/ in the code repo)| path | what | evidence |
|---|---|---|
runs/ladder_flow_*_gdag2h/policy.pt + runs/ladder_rz_*_gendag3_noqd/representation.pt | arm latent routes: flow (system i) + system 0 / representation. Lineage codes (see research/naming.md): jointfix = frozen sem s1; sejf2 = frozen sem s2; sfjf / sfjf2 = semfix (bounded probe NLL) s1 / s2; nsjf / nsjf2 = nosem s1 / s2; nszn / nsqd / nszq = nosem recipe ablation | D-078, D-091, D-095, D-099; grasp_v2 re-eval D-127 |
runs/baselines_bc_ckpts/{direct,codec}1701_u*.pt | plain BC baselines (direct action and action codec), seed 1701, with snapshots | D-064, D-083 |
runs/armexpert_bcv2/.../seed{1701,1702} | v2-teacher BC expert (DAgger labeller for the v2 set) | D-102 |
runs/armexpert_bcv6/.../seed{1701,1702} | v6 BC expert (teacher v2 + grasp_v2.1 + phase-gated DART), the DAgger labeller of the v6 lineages (1701) and the BC baseline of the target-body runs (1701, 1702). 1701 dev/fresh: panda_pg2 29/30 + 29/30, parm6_tf3 30/30 + 30/30 | D-121, D-127 |
runs/armv6/arm6-{semfix,nosem}/<node>_s{1,2} | arm v6 latent lineages (v6dart data, v6 BC labeller, grasp_v2.1), every trained node of the chain: train_rep (stage A system 0) → train_flow (F0) → flow_ft-ft → refit-bcdag1, refit-bcdag1_long, refit-bcdag2 → refit-gendag{1,2,3}_noqd → flow_ft-gdag1, flow_ft-gdag2h. Deployable route = flow_ft-gdag2h/policy.pt → refit-gendag3_noqd/representation.pt. R2 pooled semfix 425/480, nosem 314/480 (dev-seed summaries in eval_r2-final_s*/, heldout-final_s*/). The semfix train_rep, refit-bcdag1_long, refit-bcdag2 nodes are the pinned prefix of the T9 relations runs | D-134 |
runs/armtgt/armtgt6-{semfix,nosem}-<target>/target_adapt-{flow_sft,system0_refit}_b{5,20,100}_s{1,2} | v6 latent routes adapted to the target bodies xarm7_pg2, xarm7_tf3, panda_tf3 on 5/20/100 target demos (demo seeds 1,000,000+), 600 updates | D-135 |
runs/armtgt/armtgt6-bc{1701,1702}-<target>/target_adapt-bc_sft_b{5,20,100}_s* | v6 BC expert fine-tuned on the same target demos and updates (bc: source) | D-135 |
runs/armdiag/armja136-{semfix,nosem}-xarm7_{pg2,tf3}/target_adapt-joint_adapt_b{5,20,100}_s{1,2} | joint flow + system-0 adaptation on xarm7, BC-update-matched (method added AFTER D-135, pre-registered in D-136) | D-136 |
runs/legged8/legged8-{anymal_c,go2}-{semfix,nosem}/train_{rep,flow}_s{0,1,2} | legged latent routes on contact v2 | D-105, D-113, D-124 |
runs/legged8/legged8-t1sl-* | t1 humanoid on the sourced-limit tracker. CAVEAT: dataset fails the D-112 slip gate; learned controllers mostly fall | D-113, D-124 |
trackers/<body>/…/actor.pt | PPO gait trackers (contact v1 and v2, sourced-limit variants); see artifacts/trackers/README.md in the code repo for which need RRP_ACTUATOR_LIMITS=legacy_gains_v0 | D-101, D-103, D-107 |
MANIFEST.tsv lists every file with its size and sha256; rows added on 2026-10-04 also carry the source label, run id, config hash and recipe. Target-body evaluation outputs (sealed scenes), target demo packs and DAgger buffers are not included. Most checkpoints carry W3 provenance (physics/contact/grasp version,
weight fingerprint, flags) in their sidecar JSON or inside the checkpoint.
rollouts/viz_replays/ holds 160 recorded evaluation episodes (arm, dual-arm, legged) in schema rrp-viz/replay/v1 (+ v1.2 signals),
the data behind the project's visualization room: joint/body states, contacts, packet summaries and outcome labels per step.
Each episode's source (learned / bc / teacher / scripted) is labelled in its header; index.json lists them with the git sha.
import torch
ckpt = torch.load("runs/ladder_flow_sfjf2_gdag2h/policy.pt", map_location="cpu", weights_only=False)
Use the loaders in the code repo (rrp.core, rrp.controllers.bundles) at the git revision recorded in the checkpoint's
provenance. Compatibility IDs (weight fingerprints) are checked on load.
No license has been chosen yet: the code repository has no LICENSE file, so all rights are reserved by the owner until one is added. Simulated robot models come from MuJoCo Menagerie under their own licenses; no Menagerie assets are included here. Results depend on simulator versions and physics settings recorded in provenance; read the decisions log before citing numbers.
rrp checkpoints: structured latent packets for cross-body robot control
0
4 commits
1 linked in READMEs
updated Oct 4, 2026
Research checkpoints from rrp. Code, decisions and evidence live at
JacobFV/structured-psi0-latent-diffusion-dynamics
(local name relational-robot-policy). Architecture: system i is a flow model that generates a structured latent packet
z[knots × assemblies × 64], supervised by semantic probes. System 0 (our realizer) turns z plus morphology and
proprioception into joint commands; legged bodies also use a learned gait tracker. "System 0" is unrelated to the upstream Ψ₀
model (see docs/related_repos.md in the code repo).
Everything here was trained in MuJoCo simulation by this project. These are research artifacts, not deployable
controllers: no real robot has run them. Sources are labelled exactly as in the decisions log: learned:<ckpt> for our
models, bc:<ckpt> for behaviour-cloning baselines, learned_tracker:<body>:<ver> for gait trackers. Scripted teachers are not
included (they are code).
artifacts/ in the code repo)| path | what | evidence |
|---|---|---|
runs/ladder_flow_*_gdag2h/policy.pt + runs/ladder_rz_*_gendag3_noqd/representation.pt | arm latent routes: flow (system i) + system 0 / representation. Lineage codes (see research/naming.md): jointfix = frozen sem s1; sejf2 = frozen sem s2; sfjf / sfjf2 = semfix (bounded probe NLL) s1 / s2; nsjf / nsjf2 = nosem s1 / s2; nszn / nsqd / nszq = nosem recipe ablation | D-078, D-091, D-095, D-099; grasp_v2 re-eval D-127 |
runs/baselines_bc_ckpts/{direct,codec}1701_u*.pt | plain BC baselines (direct action and action codec), seed 1701, with snapshots | D-064, D-083 |
runs/armexpert_bcv2/.../seed{1701,1702} | v2-teacher BC expert (DAgger labeller for the v2 set) | D-102 |
runs/armexpert_bcv6/.../seed{1701,1702} | v6 BC expert (teacher v2 + grasp_v2.1 + phase-gated DART), the DAgger labeller of the v6 lineages (1701) and the BC baseline of the target-body runs (1701, 1702). 1701 dev/fresh: panda_pg2 29/30 + 29/30, parm6_tf3 30/30 + 30/30 | D-121, D-127 |
runs/armv6/arm6-{semfix,nosem}/<node>_s{1,2} | arm v6 latent lineages (v6dart data, v6 BC labeller, grasp_v2.1), every trained node of the chain: train_rep (stage A system 0) → train_flow (F0) → flow_ft-ft → refit-bcdag1, refit-bcdag1_long, refit-bcdag2 → refit-gendag{1,2,3}_noqd → flow_ft-gdag1, flow_ft-gdag2h. Deployable route = flow_ft-gdag2h/policy.pt → refit-gendag3_noqd/representation.pt. R2 pooled semfix 425/480, nosem 314/480 (dev-seed summaries in eval_r2-final_s*/, heldout-final_s*/). The semfix train_rep, refit-bcdag1_long, refit-bcdag2 nodes are the pinned prefix of the T9 relations runs | D-134 |
runs/armtgt/armtgt6-{semfix,nosem}-<target>/target_adapt-{flow_sft,system0_refit}_b{5,20,100}_s{1,2} | v6 latent routes adapted to the target bodies xarm7_pg2, xarm7_tf3, panda_tf3 on 5/20/100 target demos (demo seeds 1,000,000+), 600 updates | D-135 |
runs/armtgt/armtgt6-bc{1701,1702}-<target>/target_adapt-bc_sft_b{5,20,100}_s* | v6 BC expert fine-tuned on the same target demos and updates (bc: source) | D-135 |
runs/armdiag/armja136-{semfix,nosem}-xarm7_{pg2,tf3}/target_adapt-joint_adapt_b{5,20,100}_s{1,2} | joint flow + system-0 adaptation on xarm7, BC-update-matched (method added AFTER D-135, pre-registered in D-136) | D-136 |
runs/legged8/legged8-{anymal_c,go2}-{semfix,nosem}/train_{rep,flow}_s{0,1,2} | legged latent routes on contact v2 | D-105, D-113, D-124 |
runs/legged8/legged8-t1sl-* | t1 humanoid on the sourced-limit tracker. CAVEAT: dataset fails the D-112 slip gate; learned controllers mostly fall | D-113, D-124 |
trackers/<body>/…/actor.pt | PPO gait trackers (contact v1 and v2, sourced-limit variants); see artifacts/trackers/README.md in the code repo for which need RRP_ACTUATOR_LIMITS=legacy_gains_v0 | D-101, D-103, D-107 |
MANIFEST.tsv lists every file with its size and sha256; rows added on 2026-10-04 also carry the source label, run id, config hash and recipe. Target-body evaluation outputs (sealed scenes), target demo packs and DAgger buffers are not included. Most checkpoints carry W3 provenance (physics/contact/grasp version,
weight fingerprint, flags) in their sidecar JSON or inside the checkpoint.
rollouts/viz_replays/ holds 160 recorded evaluation episodes (arm, dual-arm, legged) in schema rrp-viz/replay/v1 (+ v1.2 signals),
the data behind the project's visualization room: joint/body states, contacts, packet summaries and outcome labels per step.
Each episode's source (learned / bc / teacher / scripted) is labelled in its header; index.json lists them with the git sha.
import torch
ckpt = torch.load("runs/ladder_flow_sfjf2_gdag2h/policy.pt", map_location="cpu", weights_only=False)
Use the loaders in the code repo (rrp.core, rrp.controllers.bundles) at the git revision recorded in the checkpoint's
provenance. Compatibility IDs (weight fingerprints) are checked on load.
No license has been chosen yet: the code repository has no LICENSE file, so all rights are reserved by the owner until one is added. Simulated robot models come from MuJoCo Menagerie under their own licenses; no Menagerie assets are included here. Results depend on simulator versions and physics settings recorded in provenance; read the decisions log before citing numbers.