feat(robocasa_dc): cross-embodiment benchmark + reflex-dual model servers

[MilkClouds]

What

Adds the RoboCasa-DC cross-embodiment benchmark (SeeTraceAct, arXiv:2606.02745) and its in-process reflex model servers.

• Benchmark (benchmarks/robocasa_dc): 24 RoboCasa kitchen tasks, cross-embodiment (GR-1 humanoid demo conditions a PandaOmron rollout). Scene restore is byte-identical to SeeTraceAct's validated pipeline — its restore wrapper + eval.py success criteria are imported as a module, not shelled out. Obs (53-d state + 3 flipped cameras) are pre-transformed exactly as SeeTraceAct's prepare_single_observation.
• Model servers (in-process PredictModelServer, batched + sharded GPU inference):
  • reflex_dual_robocasa_dc — dual system (S2 cognition + GR00T-N1.5 S1), GR-1 demo conditioning, per-episode cognition cache, cognition_off ablation.
  • reflex_dual_robocasa_dc_s1 — S1-only base policy (K=0 cognition).
    Both load the reflex model in-process (launched from the model repo's venv, which provides reflex) — GR00T from reflex.models.gr00t_n15, the S1-only model from reflex.models.s1.
• docker/Dockerfile.robocasa_dc for the osmesa sim deps.

Notes

• Config paths are relative / env-overridable placeholders (${oc.env:ROBOCASA_DC_EVAL_*}, /path/to/...) — no cluster-specific paths.
• The servers import reflex (the model repo) by design; they are launched from that repo's venv, so they are not exercised by this repo's own CI (deferred imports keep module import clean).

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request introduces the RoboCasa-DC cross-embodiment evaluation benchmark, including configuration files, a Dockerfile, a benchmark implementation, and model servers for both the dual-VLA and S1-only base policy. The review feedback highlights several opportunities to improve code robustness and resource management, such as replacing assertions with explicit ValueError exceptions for runtime validation, checking directory existence before listing contents, and using with open(...) context managers to prevent file descriptor leaks when loading configuration files.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

Using assert statements for runtime data validation is discouraged because they can be globally disabled in Python when run with optimization flags (e.g., python -O). It is safer to raise a ValueError instead.

Suggested change

	assert state.shape[0] == STATE_DIM, f"assembled state {state.shape[0]}d, expected {STATE_DIM}"
	if state.shape[0] != STATE_DIM:
	raise ValueError(f"assembled state {state.shape[0]}d, expected {STATE_DIM}")

[gemini-code-assist]

If the base directory does not exist, os.listdir(base) will raise a FileNotFoundError before the custom check if not seeds: is reached. It is cleaner and more robust to check if the directory exists first and raise a descriptive error.

Suggested change

	base = os.path.join(self.predefined_envs_root, task_name)
	seeds = sorted(int(d) for d in os.listdir(base) if d.isdigit() and os.path.isdir(os.path.join(base, d)))
	if not seeds:
	raise FileNotFoundError(f"no seed dirs under {base}")
	base = os.path.join(self.predefined_envs_root, task_name)
	if not os.path.isdir(base):
	raise FileNotFoundError(f"Task directory not found: {base}")
	seeds = sorted(int(d) for d in os.listdir(base) if d.isdigit() and os.path.isdir(os.path.join(base, d)))
	if not seeds:
	raise FileNotFoundError(f"no seed dirs under {base}")

[gemini-code-assist]

The file opened via open(...) is not explicitly closed, which can lead to resource leaks. Use a with open(...) context manager to ensure the file descriptor is properly released.

        with open(os.path.join(ckpt_dir, "norm_stats.json")) as f:
            self.norm_stats = NormStats.from_dict(json.load(f))

[gemini-code-assist]

The file opened via open(...) is not explicitly closed, which can lead to resource leaks. Use a with open(...) context manager to ensure the file descriptor is properly released.

        with open(os.path.join(ckpt_dir, "norm_stats.json")) as f:
            self.norm_stats = NormStats.from_dict(json.load(f))