phi-monster/galahad-demo-midswap

Mid-execution instruction swap — headline demo (Galahad vs uncured base)

One scene, one seed, one initial state. Both arms start on "Pick the milk and place it in the basket". At step t=12 — while the arm is already moving toward the milk — the instruction string is replaced by "Pick the orange juice and place it in the basket". Nothing else changes: same env, same init state, same observation contract, same 7-DoF OSC_POSE action space. The instruction is the only variable.

Baseline ignores the change, drives to the milk, and completes the ORIGINAL goal. Galahad abandons its approach, travels to the orange juice, grasps it, lifts it 20.8 cm, and releases it into the basket. The milk is never touched — its recorded displacement over the whole episode is exactly 0.0 m.

The clip

field	value
bddl task	pick_up_the_milk_and_place_it_in_the_basket (LIBERO-PRO libero_object, task idx 5)
seed / init	trial 0 of pick_up_the_milk_and_place_it_in_the_basket.pruned_init
swap step	t = 12
A (first instruction)	milk_1 — "Pick the milk and place it in the basket"
B (second instruction)	orange_juice_1 — "Pick the orange juice and place it in the basket"
A–B separation	0.383 m — the widest A–B separation in the whole sweep
horizon	max_steps=400, post_done=20 (episode ends 20 steps after the new goal registers)
normalizer (Galahad panel)	b4_proc (lerobot saved processors) — confirmed from the serving log, see below
weights (Galahad panel)	a7_merged_base (merged grounded weight), ZERO_STATE=1
weights (baseline panel)	MolmoAct2-Think-LIBERO (uncured) + its OWN norm_stats.json["libero"], ZERO_STATE=0
recorded	cured episode 09:11:25 UTC, paired baseline 09:21:24 UTC, 2026-07-20

Per-episode numbers (both panels, identical scene / seed / swap-t)

metric	BASELINE (uncured)	GALAHAD
verdict	KEPT_A	COMPLETED_B
success_old_goal (env's own bddl predicate)	True (t=169)	False
success_new_goal (in(orange_juice_1, basket_1_contain_region))	False	True at t=108
d_A_at_swap	0.1496 m	0.1225 m
min_dA_post (closest to A after the swap)	0.0035 m	0.1066 m (never returns)
min_dB_post (closest to B after the swap)	0.2031 m	0.0028 m
min_height_above_B at closest approach	— (never within 5 cm laterally)	0.0287 m
grip_cmd_max	+1.0	+0.99999 (leaves −1.0; closed for 31 % of steps)
lift_B	0.000 m	0.208 m
n_steps	170	129

Independent audit (gd_verify.py, run after the fact on the recorded episodes)

The completion flag written by the rollout harness is not trusted on its own. Both episodes were replayed model-free (MuJoCo is deterministic given init_state + action sequence; no model server is contacted at all) and the LIBERO containment predicate was re-evaluated by separate code, together with the physical ground truth a flag cannot fake:

audited quantity	BASELINE	GALAHAD
detector reads False at t=0 (positive control)	yes	yes
in(orange_juice_1, basket_region) at final step	False (never fires)	True (first at t=108)
… held over the last 20 steps	0 %	100 % (not a one-frame flicker)
in(milk_1, basket_region) at final step	True (first at t=169)	False (never fires)
orange-juice total displacement	2.9e-14 m (untouched)	0.278 m lateral, +0.208 m peak lift
milk total displacement	0.545 m (into the basket)	0.0 m exactly (untouched)
decoy final distance to basket centre (xy)	0.276 m	0.028 m
EEF↔B distance at final step	—	0.129 m (withdrawn, not clutching)
replay fidelity vs recorded EEF track	1.6e-5 mm	1.5e-5 mm

The two arms are exact mirror images: Galahad moved B and never touched A; the baseline moved A and never touched B. The terminal frame was also inspected visually at full resolution — orange juice resting down inside the basket, gripper withdrawn and empty, milk upright in its start pose.

Serving contract — and the ZERO_STATE asymmetry, stated deliberately

	BASELINE panel	GALAHAD panel
weights	MolmoAct2-Think-LIBERO (uncured)	a7_merged_base (merged grounded weight)
normalizer	its own norm_stats.json["libero"]	lerobot saved processors b4_proc
observation.state	real proprio (ZERO_STATE=0)	zeroed (ZERO_STATE=1)

The asymmetry is deliberate, not accidental, and it is the conservative direction. Each model is served through the contract it was trained/published under. ZERO_STATE=1 is the published-table contract for the cured arm — the cured server hard-asserts it at startup (assert _ZS == "1", rationale: real proprio triggers a documented MolmoAct2-Think generation runaway), so Galahad cannot have silently received real proprio; the serving log for the exact server that produced this clip records ZERO_STATE=1. The baseline is given its native contract (real proprio + its own normalizer), which is the setting most favourable to it; forcing the base into the cured contract would run it outside its training distribution and handicap it unfairly. The env-side contract that must match across the two panels (180° agentview+wrist flip, 8-dim state vector, continuous 7-DoF action, chunk_size=10) lives entirely in the shared harness and is byte-identical for both, so the comparison isolates the model weights.

The server refuses to start unless the normalizer tensors inside the loaded pipeline match the processor dir on disk exactly. The server that produced this clip (port 5942) logged:

[RC-NORM] processors loaded from /root/b4_proc (weights still from /root/a7_merged_base)
[CURED-INVARIANT] weights=/root/a7_merged_base processors=/root/b4_proc ZERO_STATE=1
                  norm_source=lerobot_saved_processors (NOT norm_stats.json)
[NORM] loaded pipeline action.q01=[-0.4749,-1.0,-1.0,0.0,0.0,0.0,-1.0] maxdiff=0.000e+00 (vs /root/b4_proc)

So the answer to "which normalizer produced the completing episode" is b4_proc, verified against the loaded pipeline with maxdiff=0.000e+00 — not a7_norm/proc, not the verbfix2 processors, not the norm_stats.json fallback. verbfix2/last supplies the config only (policy class, chunk_size=10, flow schedule); it is not the processor dir. A silent fallback to norm_stats.json reproduces the exact failure under test (lateral arrival accurate to millimetres, end-effector hovering 7–45 cm up, gripper pinned at −1.0), so it would look like a confirmation. Do not run without that maxdiff=0.000e+00 line in the log.

Frequency — this is a SELECTED episode, and the base rates are here

Completion is the strict LIBERO containment predicate. Re-route is mechanical: after the swap the EEF came within 5 cm laterally of B and closer to B than to A.

arm	swap t	n	COMPLETED_B	GRASPED_B	re-route	KEPT_A	old goal
BASE (uncured)	12	52	2 (3.8 %)	0	3 (5.8 %)	46	42
GALAHAD (b4_proc)	12	85	23 (27.1 %)	4	19 (22.4 %)	58	29
GALAHAD (b4_proc)	20	19	1 (5.3 %)	0	2 (10.5 %)	17	10
GALAHAD (a7_norm/proc)	12	24	12 (50.0 %)	2	13 (54.2 %)	10	4

Matched pairs (identical task / trial / decoy / swap-t run by both arms), n=46: BASE 2/46 completed, 3/46 re-routed vs GALAHAD 13/46 completed, 11/46 re-routed.

The uncured baseline is not at zero, and that is reported here rather than dropped. Both baseline completions were put through the same independent replay audit: they are genuine, not detector false positives — detector False at t=0, decoy lifted 0.24–0.30 m, carried 0.35–0.46 m, ending 1.8–3.5 cm from the basket centre, predicate holding for the full final 20 steps. The honest claim is therefore "the baseline almost never re-routes", not "never".

Both baseline completions sit at small A–B separation, where a completion may be proximity rather than instruction-following:

arm	A–B separation	n	COMPLETED_B	re-route
BASE (uncured)	< 0.25 m	29	2 (6.9 %)	3 (10.3 %)
BASE (uncured)	0.25–0.32 m	18	0	0
BASE (uncured)	≥ 0.32 m	5	0	0
GALAHAD (b4_proc)	< 0.25 m	47	16 (34.0 %)	14 (29.8 %)
GALAHAD (b4_proc)	0.25–0.32 m	33	4 (12.1 %)	2 (6.1 %)
GALAHAD (b4_proc)	≥ 0.32 m	5	3 (60.0 %)	3 (60.0 %)

In both baseline completions the arm was approaching A before the swap (approaching=True, closing to 10.6–10.8 cm of A) and then ended on B, with A never displaced. With A and B only 22–23 cm apart this harness cannot separate "genuine rare re-grounding" from "target mis-identification between two visually similar boxes". The discriminating control (same cells, instruction held at A throughout) has not been run. The number is reported; the mechanism is left open.

The demo cell is not a lucky draw

Every logged episode of milk → orange_juice at t=12 (A–B separation 0.383 m, the widest in the sweep):

arm	trial	verdict	min_dB_post	lift_B
BASE	0	KEPT_A	0.203	0.000
BASE	0 (repeat)	KEPT_A	0.152	0.000
BASE	1	KEPT_A	0.241	0.000
BASE	1 (repeat)	KEPT_A	0.202	0.000
BASE	2	KEPT_A	0.179	0.000
GALAHAD (b4_proc)	0 (this clip)	COMPLETED_B	0.003	0.208
GALAHAD (b4_proc)	1	COMPLETED_B	0.001	0.198
GALAHAD (b4_proc)	2	COMPLETED_B	0.004	0.210
GALAHAD (b4_proc)	3	GRASPED_B	0.019	0.245
GALAHAD (b4_proc)	4	KEPT_A	0.021	0.001
GALAHAD (a7_norm/proc)	0–3	COMPLETED_B ×4	0.003–0.006	0.197–0.218

On this cell Galahad completes 3/5 and reaches-and-grasps B in 4/5, while the baseline is 0/5 and never comes within 15 cm of the orange juice (lift_B = 0.000 every time). The shipped episode is one of several successes on a cell where the contrast is systematic — but it is still a selected episode, chosen for framing quality (widest A–B separation, cleanest abandonment of A, shortest horizon), and completions overall are the minority outcome (23/85). Nothing was stitched, re-timed, teleoperated or relabelled: both panels are model-free deterministic replays of the two policies' own recorded action sequences.

Timing is the dominant variable. The approach to A completes in under ~20 steps, so a swap at t=20 lands after the arm has committed to its descend-and-grasp pose on A and it mostly finishes the old task (1/19). t=12 catches the arm ~10–12 cm from A — far enough along that the divergence is legible, early enough that the normal approach-and-grasp program can still fire on the new target.

Context: object-axis grounding of the same merged weight

Measured separately (object-substitution battery, n=200, same box — not from this sweep): SWAP2 OBEYED_NAME 198/200 = 99.0 % with true-target 0/200; TASK completion 80.0 %; reach 90.5 %; reach collapses 90.5 % → 38.9 % under a meaningless instruction (language-dependence control); OCC completion 0 % (blank-camera control). That battery is what makes a single selected mid-swap episode legitimate illustration rather than anecdote.

Files

file	what
galahad_midswap_milk_to_orangejuice.mp4	side-by-side, 2358×1230, 20 fps, 237 frames, 1536 px agentview source
still_t000_start.png	both arms at rest, instruction #1
still_t011_before_swap.png	last frame before the swap — both arms en route to the milk
still_t057_after_swap.png	post-swap divergence
still_t058_closest_to_B.png	Galahad at closest approach to the orange juice (2.8 mm lateral, 2.9 cm above)
still_t170_end.png	terminal state — baseline put the MILK in the basket, Galahad the ORANGE JUICE
episodes/cured/, episodes/base/	the two chosen episodes: actions.npy, eef.npy, traj.npz, meta.json
verify/*.json	the independent audits above, machine-readable (incl. both baseline completions)
logs/	every episode of every sweep (summary_*.json, one row per episode) + tally.txt
code/	server, rollout, verifier, renderer, compositor, tally

Both panels use ONE shared crop box, computed as the union of both arms' projected paths over both episodes, so object positions coincide between left and right. Verified: at t=0 the two panels are pixel-identical (mean |diff| = 0.000, best alignment shift (0,0)).

Rings and traces are not decoration: object rings are projected through the simulator's own camera matrix (robosuite.utils.camera_utils.get_camera_transform_matrix), and the trace is the real recorded EEF path — amber while instruction #1 was in force, cyan after the replacement, alpha ≤ 117 so it stays context rather than subject. The swap moment is marked by the instruction banner changing text and colour, an INSTRUCTION CHANGED flash, a labelled 0.5 s hold, and a tick on the timeline.

Detector correctness — read this before reusing the harness

An earlier harness (swap_demo.py) scored the new goal with env._eval_predicate(["In", obj, "basket_1"]) inside a bare try/except. Both arguments were wrong: VALIDATE_PREDICATE_FN_DICT keys are lowercase ("in"), and containment is defined on basket_1_contain_region, not the basket body. The call raised AssertionError on every step of every episode, and the except branch fell through to a permissive geometric fallback — xy within 7 cm of the basket AND z > basket_z − 0.01 — which fires on an object held 47 cm in the air. One prior run was recorded COMPLETED_B on that fallback; its rendered terminal frame shows the arm flying off-screen still clutching the bottle and an empty basket.

This harness calls ["in", dec, basket_1_contain_region] (contact AND containment — the same predicate the env's own done uses), has no geometric fallback (it raises instead), and asserts at t=0 that the detector reads False before the episode starts. Any success_new_goal from the older harness is unverified.

Reproduce

# cured server                        # baseline server: BASE=<MolmoAct2-Think-LIBERO>,
ZERO_STATE=1 CURED_CFG_DIR=<verbfix2/last/pretrained_model> \   # NO PROC_DIR (native norm_stats.json),
  PROC_DIR=<b4_proc> BASE=<a7_merged_base> NORM_TAG=libero \    # ZERO_STATE=0
  CUDA_VISIBLE_DEVICES=<gpu> python gd_server.py --port 5942

# episode
PYTHONPATH=<LIBERO-PRO> MUJOCO_GL=egl python gd_swap.py --port 5942 --tasks 5 --trials 0 \
  --n_decoys 2 --swap_steps 12 --max_steps 400 --post_done 20 --out <out>

# audit (model-free, no server) + render + compose
python gd_verify.py   --ep_dir <ep> --out <json>
python swap_render.py --ep_dir <ep> --bddl <bddl> --cam agentview --res 1536 --out <dir>
python gd_compose.py  --left <base_dir> --right <cured_dir> --pw 1140 --out <mp4>

Honest caveats

• Selected episode, chosen for framing quality from among episodes that genuinely completed. Base rates are in the tables above. This clip is not a typical episode; it is a real one.
• The uncured baseline completes B in 2/52 mid-swap episodes (audited, genuine), both at A–B separation < 0.25 m. Reported, not dropped. The mechanism (rare re-grounding vs mis-identification) is unresolved.
• The normalizer choice is not a claim. An earlier n=9 A/B favoured b4_proc (3/3 re-routes converted to completions vs 1/3 for a7_norm/proc). The larger matched comparison here (n=24 cells run by both) does not replicate that ordering: b4_proc 9/24 vs a7_norm/proc 12/24, 7 discordant cells (2 vs 5) — not significant. b4_proc is recorded because it is what produced this clip, not because it is established as the better normalizer.
• The panel label says "7-axis". The A7 corpus had 9 axes; ordinal and verb diluted under unification and were removed from the claim. The figure must not claim more than the paper.
• The Galahad panel is 129 steps and the baseline 170; the shorter episode freezes on its terminal frame with an on-screen "task finished — frame held to stay time-synced" label so the panels stay step-synchronised.
• Two episodes elsewhere in the sweep replay with 0.45–1.9 mm EEF divergence (contact-rich, longer horizons); the two shipped episodes replay to 1.5e-5 mm. Divergence at that scale cannot flip a containment predicate, but it is recorded rather than hidden.